kiln-ai 0.17.0__py3-none-any.whl → 0.19.0__py3-none-any.whl

kiln_ai/adapters/ml_model_list.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Dict, List, Literal
+from typing import List, Literal
 
 from pydantic import BaseModel
 
@@ -33,6 +33,12 @@ class ModelFamily(str, Enum):
     deepseek = "deepseek"
     dolphin = "dolphin"
     grok = "grok"
+    kimi = "kimi"
+    hunyuan = "hunyuan"
+    glm = "glm"
+    ernie = "ernie"
+    minimax = "minimax"
+    pangu = "pangu"
 
 
 # Where models have instruct and raw versions, instruct is default and raw is specified
@@ -50,6 +56,12 @@ class ModelName(str, Enum):
     llama_3_2_11b = "llama_3_2_11b"
     llama_3_2_90b = "llama_3_2_90b"
     llama_3_3_70b = "llama_3_3_70b"
+    llama_4_maverick = "llama_4_maverick"
+    llama_4_scout = "llama_4_scout"
+    gpt_5 = "gpt_5"
+    gpt_5_chat = "gpt_5_chat"
+    gpt_5_mini = "gpt_5_mini"
+    gpt_5_nano = "gpt_5_nano"
     gpt_4o_mini = "gpt_4o_mini"
     gpt_4o = "gpt_4o"
     gpt_4_1 = "gpt_4_1"
@@ -58,6 +70,8 @@ class ModelName(str, Enum):
     gpt_o3_low = "gpt_o3_low"
     gpt_o3_medium = "gpt_o3_medium"
     gpt_o3_high = "gpt_o3_high"
+    gpt_oss_20b = "gpt_oss_20b"
+    gpt_oss_120b = "gpt_oss_120b"
     gpt_o1_low = "gpt_o1_low"
     gpt_o1_medium = "gpt_o1_medium"
     gpt_o1_high = "gpt_o1_high"
@@ -73,6 +87,9 @@ class ModelName(str, Enum):
     phi_4_mini = "phi_4_mini"
     mistral_large = "mistral_large"
     mistral_nemo = "mistral_nemo"
+    mistral_small_3 = "mistral_small_3"
+    magistral_medium = "magistral_medium"
+    magistral_medium_thinking = "magistral_medium_thinking"
     gemma_2_2b = "gemma_2_2b"
     gemma_2_9b = "gemma_2_9b"
     gemma_2_27b = "gemma_2_27b"
@@ -80,6 +97,8 @@ class ModelName(str, Enum):
     gemma_3_4b = "gemma_3_4b"
     gemma_3_12b = "gemma_3_12b"
     gemma_3_27b = "gemma_3_27b"
+    gemma_3n_2b = "gemma_3n_2b"
+    gemma_3n_4b = "gemma_3n_4b"
     claude_3_5_haiku = "claude_3_5_haiku"
     claude_3_5_sonnet = "claude_3_5_sonnet"
     claude_3_7_sonnet = "claude_3_7_sonnet"
@@ -93,6 +112,7 @@ class ModelName(str, Enum):
     gemini_2_0_flash_lite = "gemini_2_0_flash_lite"
     gemini_2_5_pro = "gemini_2_5_pro"
     gemini_2_5_flash = "gemini_2_5_flash"
+    gemini_2_5_flash_lite = "gemini_2_5_flash_lite"
     nemotron_70b = "nemotron_70b"
     mixtral_8x7b = "mixtral_8x7b"
     qwen_2p5_7b = "qwen_2p5_7b"
@@ -101,7 +121,8 @@ class ModelName(str, Enum):
     qwq_32b = "qwq_32b"
     deepseek_3 = "deepseek_3"
     deepseek_r1 = "deepseek_r1"
-    mistral_small_3 = "mistral_small_3"
+    deepseek_r1_0528 = "deepseek_r1_0528"
+    deepseek_r1_0528_distill_qwen3_8b = "deepseek_r1_0528_distill_qwen3_8b"
     deepseek_r1_distill_qwen_32b = "deepseek_r1_distill_qwen_32b"
     deepseek_r1_distill_llama_70b = "deepseek_r1_distill_llama_70b"
     deepseek_r1_distill_qwen_14b = "deepseek_r1_distill_qwen_14b"
@@ -110,6 +131,9 @@ class ModelName(str, Enum):
     deepseek_r1_distill_llama_8b = "deepseek_r1_distill_llama_8b"
     dolphin_2_9_8x22b = "dolphin_2_9_8x22b"
     grok_2 = "grok_2"
+    grok_3 = "grok_3"
+    grok_3_mini = "grok_3_mini"
+    grok_4 = "grok_4"
     qwen_3_0p6b = "qwen_3_0p6b"
     qwen_3_0p6b_no_thinking = "qwen_3_0p6b_no_thinking"
     qwen_3_1p7b = "qwen_3_1p7b"
@@ -120,12 +144,27 @@ class ModelName(str, Enum):
     qwen_3_8b_no_thinking = "qwen_3_8b_no_thinking"
     qwen_3_14b = "qwen_3_14b"
     qwen_3_14b_no_thinking = "qwen_3_14b_no_thinking"
+    qwen_3_30b_a3b_2507 = "qwen_3_30b_a3b_2507"
     qwen_3_30b_a3b = "qwen_3_30b_a3b"
+    qwen_3_30b_a3b_2507_no_thinking = "qwen_3_30b_a3b_2507_no_thinking"
     qwen_3_30b_a3b_no_thinking = "qwen_3_30b_a3b_no_thinking"
     qwen_3_32b = "qwen_3_32b"
     qwen_3_32b_no_thinking = "qwen_3_32b_no_thinking"
+    qwen_3_235b_a22b_2507 = "qwen_3_235b_a22b_2507"
     qwen_3_235b_a22b = "qwen_3_235b_a22b"
+    qwen_3_235b_a22b_2507_no_thinking = "qwen_3_235b_a22b_2507_no_thinking"
     qwen_3_235b_a22b_no_thinking = "qwen_3_235b_a22b_no_thinking"
+    qwen_long_l1_32b = "qwen_long_l1_32b"
+    kimi_k2 = "kimi_k2"
+    kimi_dev_72b = "kimi_dev_72b"
+    glm_4_1v_9b_thinking = "glm_4_1v_9b_thinking"
+    glm_z1_32b_0414 = "glm_z1_32b_0414"
+    glm_z1_9b_0414 = "glm_z1_9b_0414"
+    ernie_4_5_300b_a47b = "ernie_4_5_300b_a47b"
+    hunyuan_a13b = "hunyuan_a13b"
+    hunyuan_a13b_no_thinking = "hunyuan_a13b_no_thinking"
+    minimax_m1_80k = "minimax_m1_80k"
+    pangu_pro_moe_72b_a16b = "pangu_pro_moe_72b_a16b"
 
 
 class ModelParserID(str, Enum):
@@ -174,9 +213,11 @@ class KilnModelProvider(BaseModel):
     reasoning_capable: bool = False
     supports_logprobs: bool = False
     suggested_for_evals: bool = False
+    uncensored: bool = False
+    suggested_for_uncensored_data_gen: bool = False
     tuned_chat_strategy: ChatStrategy | None = None
 
-    # TODO P1: Need a more generalized way to handle custom provider parameters.
+    # We need a more generalized way to handle custom provider parameters.
     # Making them quite declarative here for now, isolating provider specific logic
     # to this file. Later I should be able to override anything in this file via config.
     r1_openrouter_options: bool = False
@@ -186,6 +227,16 @@ class KilnModelProvider(BaseModel):
     thinking_level: Literal["low", "medium", "high"] | None = None
     ollama_model_aliases: List[str] | None = None
     anthropic_extended_thinking: bool = False
+    gemini_reasoning_enabled: bool = False
+
+    # some models on siliconflow allow dynamically disabling thinking
+    # currently only supported by Qwen3 and tencent/Hunyuan-A13B-Instruct
+    # ref: https://docs.siliconflow.cn/cn/api-reference/chat-completions/chat-completions
+    siliconflow_enable_thinking: bool | None = None
+    # enable this flag to make reasoning optional for structured output
+    # some reasoning models on siliconflow do not return any reasoning for structured output
+    # this is not uniform nor documented, so we need to test each model
+    reasoning_optional_for_structured_output: bool | None = None
 
 
 class KilnModel(BaseModel):
@@ -207,6 +258,63 @@ class KilnModel(BaseModel):
 
 
 built_in_models: List[KilnModel] = [
+    # GPT 5
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_5,
+        friendly_name="GPT-5",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="gpt-5",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                suggested_for_data_gen=True,
+                suggested_for_evals=True,
+            ),
+        ],
+    ),
+    # GPT 5 Mini
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_5_mini,
+        friendly_name="GPT-5 Mini",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="gpt-5-mini",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                suggested_for_evals=True,
+                suggested_for_data_gen=True,
+            ),
+        ],
+    ),
+    # GPT 5 Nano
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_5_nano,
+        friendly_name="GPT-5 Nano",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="gpt-5-nano",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+        ],
+    ),
+    # GPT 5 Chat
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_5_chat,
+        friendly_name="GPT-5 Chat",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="gpt-5-chat-latest",
+                # Oddly no json_schema support for this model.
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+        ],
+    ),
     # GPT 4.1
     KilnModel(
         family=ModelFamily.gpt,
@@ -219,7 +327,6 @@ built_in_models: List[KilnModel] = [
                 provider_finetune_id="gpt-4.1-2025-04-14",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
-                suggested_for_data_gen=True,
                 suggested_for_evals=True,
             ),
             KilnModelProvider(
@@ -227,13 +334,11 @@ built_in_models: List[KilnModel] = [
                 model_id="openai/gpt-4.1",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
-                suggested_for_data_gen=True,
                 suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.azure_openai,
                 model_id="gpt-4.1",
-                suggested_for_data_gen=True,
                 suggested_for_evals=True,
             ),
         ],
@@ -250,22 +355,16 @@ built_in_models: List[KilnModel] = [
                 provider_finetune_id="gpt-4.1-mini-2025-04-14",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="openai/gpt-4.1-mini",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.azure_openai,
                 model_id="gpt-4.1-mini",
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
         ],
     ),
@@ -280,6 +379,7 @@ built_in_models: List[KilnModel] = [
                 model_id="gpt-4.1-nano",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
+                provider_finetune_id="gpt-4.1-nano-2025-04-14",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
@@ -305,6 +405,8 @@ built_in_models: List[KilnModel] = [
                 provider_finetune_id="gpt-4o-2024-08-06",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
+                suggested_for_data_gen=True,
+                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
@@ -312,10 +414,14 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_logprobs=True,
                 logprobs_openrouter_options=True,
+                suggested_for_data_gen=True,
+                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.azure_openai,
                 model_id="gpt-4o",
+                suggested_for_data_gen=True,
+                suggested_for_evals=True,
             ),
         ],
     ),
@@ -535,6 +641,70 @@ built_in_models: List[KilnModel] = [
             ),
         ],
     ),
+    # GPT OSS 120B
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_oss_120b,
+        friendly_name="GPT OSS 120B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="openai/gpt-oss-120b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                require_openrouter_reasoning=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.groq,
+                model_id="openai/gpt-oss-120b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/gpt-oss-120b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gpt-oss:120b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+        ],
+    ),
+    # GPT OSS 20B
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_oss_20b,
+        friendly_name="GPT OSS 20B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="openai/gpt-oss-20b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                require_openrouter_reasoning=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.groq,
+                model_id="openai/gpt-oss-20b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/gpt-oss-20b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gpt-oss:20b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+        ],
+    ),
     # GPT o1 Low
     KilnModel(
         family=ModelFamily.gpt,
@@ -651,13 +821,11 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.function_calling,
                 model_id="anthropic/claude-3.7-sonnet",
-                suggested_for_data_gen=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.anthropic,
                 model_id="claude-3-7-sonnet-20250219",
                 structured_output_mode=StructuredOutputMode.function_calling,
-                suggested_for_data_gen=True,
             ),
         ],
     ),
@@ -732,24 +900,33 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                model_id="google/gemini-2.5-pro-preview-03-25",
+                model_id="google/gemini-2.5-pro",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 suggested_for_data_gen=True,
                 suggested_for_evals=True,
+                reasoning_capable=True,
+                gemini_reasoning_enabled=True,
+                thinking_level="medium",
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
-                model_id="gemini-2.5-pro-preview-03-25",
+                model_id="gemini-2.5-pro",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 suggested_for_data_gen=True,
                 suggested_for_evals=True,
+                reasoning_capable=True,
+                gemini_reasoning_enabled=True,
+                thinking_level="medium",
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
-                model_id="gemini-2.5-pro-preview-03-25",
+                model_id="gemini-2.5-pro",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 suggested_for_data_gen=True,
                 suggested_for_evals=True,
+                reasoning_capable=True,
+                gemini_reasoning_enabled=True,
+                thinking_level="medium",
             ),
         ],
     ),
@@ -761,18 +938,24 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                model_id="google/gemini-2.5-flash-preview",
+                model_id="google/gemini-2.5-flash",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                gemini_reasoning_enabled=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
-                model_id="gemini-2.5-flash-preview-04-17",
+                model_id="gemini-2.5-flash",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                thinking_level="medium",
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
-                model_id="gemini-2.5-flash-preview-04-17",
+                model_id="gemini-2.5-flash",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                thinking_level="medium",
             ),
         ],
     ),
@@ -785,17 +968,17 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="google/gemini-2.0-flash-001",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
                 model_id="gemini-2.0-flash",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
                 model_id="gemini-2.0-flash",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
                 provider_finetune_id="gemini-2.0-flash-001",
             ),
         ],
@@ -809,17 +992,17 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="google/gemini-2.0-flash-lite-001",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
                 model_id="gemini-2.0-flash-lite",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
                 model_id="gemini-2.0-flash-lite",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
                 provider_finetune_id="gemini-2.0-flash-lite-001",
             ),
         ],
@@ -833,12 +1016,12 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="google/gemini-pro-1.5",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
                 model_id="gemini-1.5-pro",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
@@ -856,12 +1039,12 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="google/gemini-flash-1.5",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
                 model_id="gemini-1.5-flash",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
             KilnModelProvider(
                 name=ModelProviderName.vertex,
@@ -879,13 +1062,13 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="google/gemini-flash-1.5-8b",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
                 supports_data_gen=False,
             ),
             KilnModelProvider(
                 name=ModelProviderName.gemini_api,
                 model_id="gemini-1.5-flash-8b",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                structured_output_mode=StructuredOutputMode.json_schema,
                 supports_data_gen=False,
             ),
         ],
@@ -904,6 +1087,62 @@ built_in_models: List[KilnModel] = [
             ),
         ],
     ),
+    # Llama 4 Maverick Basic
+    KilnModel(
+        family=ModelFamily.llama,
+        name=ModelName.llama_4_maverick,
+        friendly_name="Llama 4 Maverick",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="meta-llama/llama-4-maverick",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/llama4-maverick-instruct-basic",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.cerebras,
+                model_id="llama-4-maverick-17b-128e-instruct",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+        ],
+    ),
+    # Llama 4 Scout Basic
+    KilnModel(
+        family=ModelFamily.llama,
+        name=ModelName.llama_4_scout,
+        friendly_name="Llama 4 Scout",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="meta-llama/llama-4-scout",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/llama4-scout-instruct-basic",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Llama-4-Scout-17B-16E-Instruct",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.cerebras,
+                model_id="llama-4-scout-17b-16e-instruct",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+        ],
+    ),
     # Llama 3.1-8b
     KilnModel(
         family=ModelFamily.llama,
@@ -946,6 +1185,13 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.function_calling_weak,
                 provider_finetune_id="meta-llama/Meta-Llama-3.1-8B-Instruct-Reference",
             ),
+            KilnModelProvider(
+                name=ModelProviderName.cerebras,
+                model_id="llama3.1-8b",
+                structured_output_mode=StructuredOutputMode.function_calling,
+                supports_data_gen=False,
+                suggested_for_evals=False,
+            ),
         ],
     ),
     # Llama 3.1 70b
@@ -1024,6 +1270,33 @@ built_in_models: List[KilnModel] = [
             ),
         ],
     ),
+    # Magistral Medium (Thinking)
+    KilnModel(
+        family=ModelFamily.mistral,
+        name=ModelName.magistral_medium_thinking,
+        friendly_name="Magistral Medium (Thinking)",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="mistralai/magistral-medium-2506:thinking",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                # Thinking tokens are hidden by Mistral so not "reasoning" from Kiln API POV
+            ),
+        ],
+    ),
+    # Magistral Medium (No Thinking)
+    KilnModel(
+        family=ModelFamily.mistral,
+        name=ModelName.magistral_medium,
+        friendly_name="Magistral Medium (No Thinking)",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="mistralai/magistral-medium-2506",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+        ],
+    ),
     # Mistral Nemo
     KilnModel(
         family=ModelFamily.mistral,
@@ -1047,16 +1320,22 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.amazon_bedrock,
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 model_id="mistral.mistral-large-2407-v1:0",
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_schema,
                 model_id="mistralai/mistral-large",
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
                 model_id="mistral-large",
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
             ),
         ],
     ),
@@ -1066,12 +1345,6 @@ built_in_models: List[KilnModel] = [
         name=ModelName.llama_3_2_1b,
         friendly_name="Llama 3.2 1B",
         providers=[
-            KilnModelProvider(
-                name=ModelProviderName.groq,
-                model_id="llama-3.2-1b-preview",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                supports_data_gen=False,
-            ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_structured_output=False,
@@ -1085,12 +1358,6 @@ built_in_models: List[KilnModel] = [
                 supports_data_gen=False,
                 model_id="llama3.2:1b",
             ),
-            KilnModelProvider(
-                name=ModelProviderName.huggingface,
-                model_id="meta-llama/Llama-3.2-1B-Instruct",
-                supports_structured_output=False,
-                supports_data_gen=False,
-            ),
         ],
     ),
     # Llama 3.2 3B
@@ -1099,11 +1366,6 @@ built_in_models: List[KilnModel] = [
         name=ModelName.llama_3_2_3b,
         friendly_name="Llama 3.2 3B",
         providers=[
-            KilnModelProvider(
-                name=ModelProviderName.groq,
-                model_id="llama-3.2-3b-preview",
-                supports_data_gen=False,
-            ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_structured_output=False,
@@ -1116,12 +1378,6 @@ built_in_models: List[KilnModel] = [
                 supports_data_gen=False,
                 model_id="llama3.2",
             ),
-            KilnModelProvider(
-                name=ModelProviderName.huggingface,
-                model_id="meta-llama/Llama-3.2-3B-Instruct",
-                supports_structured_output=False,
-                supports_data_gen=False,
-            ),
             KilnModelProvider(
                 name=ModelProviderName.together_ai,
                 model_id="meta-llama/Llama-3.2-3B-Instruct-Turbo",
@@ -1149,19 +1405,6 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_schema,
                 model_id="llama3.2-vision",
             ),
-            KilnModelProvider(
-                name=ModelProviderName.fireworks_ai,
-                # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                model_id="accounts/fireworks/models/llama-v3p2-11b-vision-instruct",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                supports_data_gen=False,
-            ),
-            KilnModelProvider(
-                name=ModelProviderName.huggingface,
-                model_id="meta-llama/Llama-3.2-11B-Vision-Instruct",
-                supports_structured_output=False,
-                supports_data_gen=False,
-            ),
             KilnModelProvider(
                 name=ModelProviderName.together_ai,
                 model_id="meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo",
@@ -1186,13 +1429,6 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_schema,
                 model_id="llama3.2-vision:90b",
             ),
-            KilnModelProvider(
-                name=ModelProviderName.fireworks_ai,
-                # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                model_id="accounts/fireworks/models/llama-v3p2-90b-vision-instruct",
-                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                supports_data_gen=False,
-            ),
             KilnModelProvider(
                 name=ModelProviderName.together_ai,
                 model_id="meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo",
@@ -1235,7 +1471,7 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.vertex,
                 model_id="meta/llama-3.3-70b-instruct-maas",
-                # Doesn't work, TODO to debug
+                # Doesn't work yet; needs debugging
                 supports_structured_output=False,
                 supports_data_gen=False,
             ),
@@ -1266,13 +1502,6 @@ built_in_models: List[KilnModel] = [
                 model_id="microsoft/phi-3.5-mini-128k-instruct",
                 structured_output_mode=StructuredOutputMode.json_schema,
             ),
-            KilnModelProvider(
-                name=ModelProviderName.fireworks_ai,
-                # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                supports_structured_output=False,
-                supports_data_gen=False,
-                model_id="accounts/fireworks/models/phi-3-vision-128k-instruct",
-            ),
         ],
     ),
     # Phi 4
@@ -1387,13 +1616,6 @@ built_in_models: List[KilnModel] = [
                 supports_structured_output=False,
                 supports_data_gen=False,
             ),
-            KilnModelProvider(
-                name=ModelProviderName.openrouter,
-                # TODO: swap to non-free model when available (more reliable)
-                model_id="google/gemma-3-1b-it:free",
-                supports_structured_output=False,
-                supports_data_gen=False,
-            ),
         ],
     ),
     # Gemma 3 4B
@@ -1410,8 +1632,7 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                # TODO: swap to non-free model when available (more reliable)
-                model_id="google/gemma-3-4b-it:free",
+                model_id="google/gemma-3-4b-it",
             ),
         ],
     ),
@@ -1428,8 +1649,7 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                # TODO: swap to non-free model when available (more reliable)
-                model_id="google/gemma-3-12b-it:free",
+                model_id="google/gemma-3-12b-it",
             ),
         ],
     ),
@@ -1450,11 +1670,57 @@ built_in_models: List[KilnModel] = [
             ),
         ],
     ),
-    # Mixtral 8x7B
+    # Gemma 3n 2B
     KilnModel(
-        family=ModelFamily.mixtral,
-        name=ModelName.mixtral_8x7b,
-        friendly_name="Mixtral 8x7B",
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3n_2b,
+        friendly_name="Gemma 3n 2B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3n:e2b",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemma-3n-e2b-it",
+                supports_structured_output=False,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Gemma 3n 4B
+    KilnModel(
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3n_4b,
+        friendly_name="Gemma 3n 4B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="google/gemma-3n-e4b-it",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_data_gen=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3n:e4b",
+                supports_data_gen=False,
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemma-3n-e4b-it",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Mixtral 8x7B
+    KilnModel(
+        family=ModelFamily.mixtral,
+        name=ModelName.mixtral_8x7b,
+        friendly_name="Mixtral 8x7B",
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
@@ -1498,17 +1764,16 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
             ),
             KilnModelProvider(
-                name=ModelProviderName.groq,
-                model_id="qwen-qwq-32b",
-                reasoning_capable=True,
-                parser=ModelParserID.r1_thinking,
+                name=ModelProviderName.together_ai,
+                model_id="Qwen/QwQ-32B",
                 structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+                reasoning_capable=True,
             ),
             KilnModelProvider(
-                name=ModelProviderName.together_ai,
+                name=ModelProviderName.siliconflow_cn,
                 model_id="Qwen/QwQ-32B",
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
             ),
         ],
@@ -1587,10 +1852,78 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
                 model_id="mistralai/mistral-small-24b-instruct-2501",
+                uncensored=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 model_id="mistral-small:24b",
+                uncensored=True,
+            ),
+        ],
+    ),
+    # DeepSeek R1 0528
+    KilnModel(
+        family=ModelFamily.deepseek,
+        name=ModelName.deepseek_r1_0528,
+        friendly_name="DeepSeek R1 0528",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="deepseek/deepseek-r1-0528",
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+                supports_data_gen=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/deepseek-r1-0528",
+                parser=ModelParserID.r1_thinking,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="deepseek-ai/DeepSeek-R1",  # Note: Together remapped the R1 endpoint to this 0528 model
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+                reasoning_capable=True,
+                supports_data_gen=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Pro/deepseek-ai/DeepSeek-R1",
+                parser=ModelParserID.optional_r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # DeepSeek R1 0528 Distill Qwen 3 8B
+    KilnModel(
+        family=ModelFamily.deepseek,
+        name=ModelName.deepseek_r1_0528_distill_qwen3_8b,
+        friendly_name="DeepSeek R1 0528 Distill Qwen 3 8B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="deepseek/deepseek-r1-0528-qwen3-8b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="deepseek-ai/DeepSeek-R1-0528-Qwen3-8B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
             ),
         ],
     ),
@@ -1617,13 +1950,18 @@ built_in_models: List[KilnModel] = [
                 model_id="deepseek-ai/DeepSeek-V3",
                 structured_output_mode=StructuredOutputMode.json_instructions,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Pro/deepseek-ai/DeepSeek-V3",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
         ],
     ),
     # DeepSeek R1
     KilnModel(
         family=ModelFamily.deepseek,
         name=ModelName.deepseek_r1,
-        friendly_name="DeepSeek R1",
+        friendly_name="DeepSeek R1 (Original)",
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
@@ -1633,8 +1971,6 @@ built_in_models: List[KilnModel] = [
                 reasoning_capable=True,
                 r1_openrouter_options=True,
                 require_openrouter_reasoning=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
@@ -1642,8 +1978,6 @@ built_in_models: List[KilnModel] = [
                 parser=ModelParserID.r1_thinking,
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 reasoning_capable=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
             KilnModelProvider(
                 # I want your RAM
@@ -1652,17 +1986,6 @@ built_in_models: List[KilnModel] = [
                 parser=ModelParserID.r1_thinking,
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 reasoning_capable=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
-            ),
-            KilnModelProvider(
-                name=ModelProviderName.together_ai,
-                model_id="deepseek-ai/DeepSeek-R1",
-                structured_output_mode=StructuredOutputMode.json_instructions,
-                parser=ModelParserID.r1_thinking,
-                reasoning_capable=True,
-                suggested_for_data_gen=True,
-                suggested_for_evals=True,
             ),
         ],
     ),
@@ -1689,11 +2012,11 @@ built_in_models: List[KilnModel] = [
                 model_id="deepseek-r1:32b",
             ),
             KilnModelProvider(
-                name=ModelProviderName.together_ai,
-                model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
-                structured_output_mode=StructuredOutputMode.json_instructions,
-                parser=ModelParserID.r1_thinking,
+                name=ModelProviderName.siliconflow_cn,
+                model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
+                structured_output_mode=StructuredOutputMode.json_schema,
                 reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
             ),
         ],
     ),
@@ -1759,6 +2082,14 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 parser=ModelParserID.r1_thinking,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
+            ),
         ],
     ),
     # DeepSeek R1 Distill Llama 8B
@@ -1798,6 +2129,17 @@ built_in_models: List[KilnModel] = [
         name=ModelName.deepseek_r1_distill_qwen_7b,
         friendly_name="DeepSeek R1 Distill Qwen 7B",
         providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                # Best mode, but fails to often to enable without warning
+                supports_structured_output=False,
+                supports_data_gen=False,
+                model_id="deepseek/deepseek-r1-distill-qwen-7b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+            ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 # Best mode, but fails to often to enable without warning
@@ -1808,6 +2150,16 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 model_id="deepseek-r1:7b",
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                # Best mode, but fails to often to enable without warning
+                supports_structured_output=False,
+                supports_data_gen=False,
+                model_id="Pro/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+            ),
         ],
     ),
     # DeepSeek R1 Distill Qwen 1.5B
@@ -1858,12 +2210,67 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_data_gen=True,
                 model_id="dolphin-mixtral:8x22b",
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_data_gen=True,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
                 model_id="cognitivecomputations/dolphin-mixtral-8x22b",
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
+            ),
+        ],
+    ),
+    # Grok 4
+    KilnModel(
+        family=ModelFamily.grok,
+        name=ModelName.grok_4,
+        friendly_name="Grok 4",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="x-ai/grok-4",
+                supports_structured_output=True,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
+                suggested_for_data_gen=True,
+                uncensored=True,
+                suggested_for_uncensored_data_gen=True,
+            ),
+        ],
+    ),
+    # Grok 3
+    KilnModel(
+        family=ModelFamily.grok,
+        name=ModelName.grok_3,
+        friendly_name="Grok 3",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="x-ai/grok-3",
+                supports_structured_output=True,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
+                suggested_for_data_gen=True,
+                uncensored=True,
+            ),
+        ],
+    ),
+    # Grok 3 Mini
+    KilnModel(
+        family=ModelFamily.grok,
+        name=ModelName.grok_3_mini,
+        friendly_name="Grok 3 Mini",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="x-ai/grok-3-mini",
+                supports_structured_output=True,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
+                uncensored=True,
             ),
         ],
     ),
@@ -2028,6 +2435,15 @@ built_in_models: List[KilnModel] = [
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_schema,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-8B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                siliconflow_enable_thinking=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
+            ),
         ],
     ),
     # Qwen 3 8B Non-Thinking
@@ -2051,6 +2467,13 @@ built_in_models: List[KilnModel] = [
                 formatter=ModelFormatterID.qwen3_style_no_think,
                 supports_data_gen=False,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-8B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                siliconflow_enable_thinking=False,
+                supports_data_gen=False,
+            ),
         ],
     ),
     # Qwen 3 14B
@@ -2076,6 +2499,15 @@ built_in_models: List[KilnModel] = [
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_schema,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-14B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+                reasoning_capable=True,
+                siliconflow_enable_thinking=True,
+                reasoning_optional_for_structured_output=True,
+            ),
         ],
     ),
     # Qwen 3 14B Non-Thinking
@@ -2099,6 +2531,28 @@ built_in_models: List[KilnModel] = [
                 supports_data_gen=True,
                 structured_output_mode=StructuredOutputMode.json_schema,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-14B",
+                formatter=ModelFormatterID.qwen3_style_no_think,
+                structured_output_mode=StructuredOutputMode.json_schema,
+                siliconflow_enable_thinking=False,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # Qwen 3 30B (3B Active) 2507 Version
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_3_30b_a3b_2507,
+        friendly_name="Qwen 3 30B (3B Active) 2507",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwen3:30b-a3b-thinking-2507-q4_K_M",
+                reasoning_capable=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
         ],
     ),
     # Qwen 3 30B (3B Active)
@@ -2132,6 +2586,33 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 parser=ModelParserID.r1_thinking,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-30B-A3B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # Qwen 3 30B (3B Active) 2507 Version Non-Thinking
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_3_30b_a3b_2507_no_thinking,
+        friendly_name="Qwen 3 30B (3B Active) 2507 Non-Thinking",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="qwen/qwen3-30b-a3b-instruct-2507",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwen3:30b-a3b-instruct-2507-q8_0",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+            ),
         ],
     ),
     # Qwen 3 30B (3B Active) Non-Thinking
@@ -2171,6 +2652,14 @@ built_in_models: List[KilnModel] = [
         name=ModelName.qwen_3_32b,
         friendly_name="Qwen 3 32B",
         providers=[
+            KilnModelProvider(
+                name=ModelProviderName.groq,
+                model_id="Qwen/Qwen3-32B",
+                supports_data_gen=True,
+                reasoning_capable=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+            ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 model_id="qwen/qwen3-32b",
@@ -2188,6 +2677,22 @@ built_in_models: List[KilnModel] = [
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_schema,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-32B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.cerebras,
+                model_id="qwen-3-32b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                supports_data_gen=True,
+                reasoning_capable=True,
+                parser=ModelParserID.r1_thinking,
+            ),
         ],
     ),
     # Qwen 3 32B No Thinking
@@ -2211,6 +2716,56 @@ built_in_models: List[KilnModel] = [
                 formatter=ModelFormatterID.qwen3_style_no_think,
                 supports_data_gen=True,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.cerebras,
+                model_id="qwen-3-32b",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                formatter=ModelFormatterID.qwen3_style_no_think,
+                supports_data_gen=True,
+                parser=ModelParserID.optional_r1_thinking,
+            ),
+        ],
+    ),
+    # Qwen 3 235B (22B Active) 2507 Version
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_3_235b_a22b_2507,
+        friendly_name="Qwen 3 235B (22B Active) 2507",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="qwen/qwen3-235b-a22b-thinking-2507",
+                reasoning_capable=True,
+                require_openrouter_reasoning=True,
+                supports_data_gen=True,
+                suggested_for_data_gen=True,
+                r1_openrouter_options=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwen3:235b-a22b-thinking-2507-q4_K_M",
+                supports_data_gen=True,
+                reasoning_capable=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/qwen3-235b-a22b-thinking-2507",
+                supports_data_gen=True,
+                reasoning_capable=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="Qwen/Qwen3-235B-A22B-Thinking-2507",
+                supports_data_gen=True,
+                reasoning_capable=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+            ),
         ],
     ),
     # Qwen 3 235B (22B Active)
@@ -2225,7 +2780,6 @@ built_in_models: List[KilnModel] = [
                 reasoning_capable=True,
                 require_openrouter_reasoning=True,
                 supports_data_gen=True,
-                suggested_for_data_gen=True,
                 r1_openrouter_options=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 parser=ModelParserID.r1_thinking,
@@ -2253,6 +2807,49 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 parser=ModelParserID.r1_thinking,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-235B-A22B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                siliconflow_enable_thinking=True,
+                supports_data_gen=True,
+                suggested_for_data_gen=True,
+            ),
+        ],
+    ),
+    # Qwen 3 235B (22B Active) 2507 Version Non-Thinking
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_3_235b_a22b_2507_no_thinking,
+        friendly_name="Qwen 3 235B (22B Active) 2507 Non-Thinking",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="qwen/qwen3-235b-a22b-2507",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                supports_data_gen=True,
+                reasoning_capable=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwen3:235b-a22b-instruct-2507-q4_K_M",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+                reasoning_capable=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/qwen3-235b-a22b-instruct-2507",
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="Qwen/Qwen3-235B-A22B-Instruct-2507-tput",
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
         ],
     ),
     # Qwen 3 235B (22B Active) Non-Thinking
@@ -2293,6 +2890,210 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_instructions,
                 parser=ModelParserID.optional_r1_thinking,
             ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Qwen/Qwen3-235B-A22B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                siliconflow_enable_thinking=False,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # Qwen Long L1 32B
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_long_l1_32b,
+        friendly_name="QwenLong L1 32B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Tongyi-Zhiwen/QwenLong-L1-32B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+            ),
+        ],
+    ),
+    # Kimi K2 Instruct
+    KilnModel(
+        family=ModelFamily.kimi,
+        name=ModelName.kimi_k2,
+        friendly_name="Kimi K2",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/kimi-k2-instruct",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                # Ignoring json mode for now, so not suggested for evals
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="moonshotai/kimi-k2",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+                suggested_for_evals=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="moonshotai/Kimi-K2-Instruct",
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                suggested_for_evals=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.groq,
+                model_id="moonshotai/kimi-k2-instruct",
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.function_calling,
+                suggested_for_evals=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Pro/moonshotai/Kimi-K2-Instruct",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+                suggested_for_evals=True,
+            ),
+        ],
+    ),
+    KilnModel(
+        family=ModelFamily.kimi,
+        name=ModelName.kimi_dev_72b,
+        friendly_name="Kimi Dev 72B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="moonshotai/Kimi-Dev-72B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+            ),
+        ],
+    ),
+    # GLM 4.1V 9B
+    KilnModel(
+        family=ModelFamily.glm,
+        name=ModelName.glm_4_1v_9b_thinking,
+        friendly_name="GLM-4.1V 9B Thinking",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="Pro/THUDM/GLM-4.1V-9B-Thinking",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # GLM Z1 32B 0414
+    KilnModel(
+        family=ModelFamily.glm,
+        name=ModelName.glm_z1_32b_0414,
+        friendly_name="GLM-Z1 32B 0414",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="THUDM/GLM-Z1-32B-0414",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # GLM Z1 9B 0414
+    KilnModel(
+        family=ModelFamily.glm,
+        name=ModelName.glm_z1_9b_0414,
+        friendly_name="GLM-Z1 9B 0414",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="THUDM/GLM-Z1-9B-0414",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Ernie 4.5 300B A47B
+    KilnModel(
+        family=ModelFamily.ernie,
+        name=ModelName.ernie_4_5_300b_a47b,
+        friendly_name="Ernie 4.5 300B A47B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="baidu/ernie-4.5-300b-a47b",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                supports_data_gen=True,
+                r1_openrouter_options=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="baidu/ERNIE-4.5-300B-A47B",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # Hunyuan A13B Instruct
+    KilnModel(
+        family=ModelFamily.hunyuan,
+        name=ModelName.hunyuan_a13b,
+        friendly_name="Hunyuan A13B",
+        providers=[
+            # Openrouter provider for this model exists but currently wrongly parses the answer
+            # it returns the reasoning at the right place, but wraps the answer (even JSON response)
+            # between <answer> and </answer> tags
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="tencent/Hunyuan-A13B-Instruct",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                reasoning_capable=True,
+                siliconflow_enable_thinking=True,
+                reasoning_optional_for_structured_output=True,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Minimax M1 80K
+    KilnModel(
+        family=ModelFamily.minimax,
+        name=ModelName.minimax_m1_80k,
+        friendly_name="Minimax M1",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="minimax/minimax-m1",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                supports_data_gen=True,
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="MiniMaxAI/MiniMax-M1-80k",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                supports_data_gen=True,
+            ),
+        ],
+    ),
+    # Pangu Pro MOE
+    KilnModel(
+        family=ModelFamily.pangu,
+        name=ModelName.pangu_pro_moe_72b_a16b,
+        friendly_name="Pangu Pro MOE 72B A16B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.siliconflow_cn,
+                model_id="ascend-tribe/pangu-pro-moe",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                supports_data_gen=True,
+            ),
         ],
     ),
 ]