kiln-ai 0.11.1__py3-none-any.whl → 0.13.0__py3-none-any.whl

kiln_ai/adapters/ml_model_list.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Dict, List
+from typing import Dict, List, Literal
 
 from pydantic import BaseModel
 
@@ -26,6 +26,12 @@ class ModelProviderName(str, Enum):
     kiln_fine_tune = "kiln_fine_tune"
     kiln_custom_registry = "kiln_custom_registry"
     openai_compatible = "openai_compatible"
+    anthropic = "anthropic"
+    gemini_api = "gemini_api"
+    azure_openai = "azure_openai"
+    huggingface = "huggingface"
+    vertex = "vertex"
+    together_ai = "together_ai"
 
 
 class ModelFamily(str, Enum):
@@ -43,6 +49,8 @@ class ModelFamily(str, Enum):
     mixtral = "mixtral"
     qwen = "qwen"
     deepseek = "deepseek"
+    dolphin = "dolphin"
+    grok = "grok"
 
 
 # Where models have instruct and raw versions, instruct is default and raw is specified
@@ -62,15 +70,29 @@ class ModelName(str, Enum):
     llama_3_3_70b = "llama_3_3_70b"
     gpt_4o_mini = "gpt_4o_mini"
     gpt_4o = "gpt_4o"
+    gpt_o1_low = "gpt_o1_low"
+    gpt_o1_medium = "gpt_o1_medium"
+    gpt_o1_high = "gpt_o1_high"
+    gpt_o3_mini_low = "gpt_o3_mini_low"
+    gpt_o3_mini_medium = "gpt_o3_mini_medium"
+    gpt_o3_mini_high = "gpt_o3_mini_high"
     phi_3_5 = "phi_3_5"
     phi_4 = "phi_4"
+    phi_4_5p6b = "phi_4_5p6b"
+    phi_4_mini = "phi_4_mini"
     mistral_large = "mistral_large"
     mistral_nemo = "mistral_nemo"
     gemma_2_2b = "gemma_2_2b"
     gemma_2_9b = "gemma_2_9b"
     gemma_2_27b = "gemma_2_27b"
+    gemma_3_1b = "gemma_3_1b"
+    gemma_3_4b = "gemma_3_4b"
+    gemma_3_12b = "gemma_3_12b"
+    gemma_3_27b = "gemma_3_27b"
     claude_3_5_haiku = "claude_3_5_haiku"
     claude_3_5_sonnet = "claude_3_5_sonnet"
+    claude_3_7_sonnet = "claude_3_7_sonnet"
+    claude_3_7_sonnet_thinking = "claude_3_7_sonnet_thinking"
     gemini_1_5_flash = "gemini_1_5_flash"
     gemini_1_5_flash_8b = "gemini_1_5_flash_8b"
     gemini_1_5_pro = "gemini_1_5_pro"
@@ -78,7 +100,9 @@ class ModelName(str, Enum):
     nemotron_70b = "nemotron_70b"
     mixtral_8x7b = "mixtral_8x7b"
     qwen_2p5_7b = "qwen_2p5_7b"
+    qwen_2p5_14b = "qwen_2p5_14b"
     qwen_2p5_72b = "qwen_2p5_72b"
+    qwq_32b = "qwq_32b"
     deepseek_3 = "deepseek_3"
     deepseek_r1 = "deepseek_r1"
     mistral_small_3 = "mistral_small_3"
@@ -88,6 +112,8 @@ class ModelName(str, Enum):
     deepseek_r1_distill_qwen_1p5b = "deepseek_r1_distill_qwen_1p5b"
     deepseek_r1_distill_qwen_7b = "deepseek_r1_distill_qwen_7b"
     deepseek_r1_distill_llama_8b = "deepseek_r1_distill_llama_8b"
+    dolphin_2_9_8x22b = "dolphin_2_9_8x22b"
+    grok_2 = "grok_2"
 
 
 class ModelParserID(str, Enum):
@@ -108,21 +134,32 @@ class KilnModelProvider(BaseModel):
         supports_data_gen: Whether the provider supports data generation
         untested_model: Whether the model is untested (typically user added). The supports_ fields are not applicable.
         provider_finetune_id: The finetune ID for the provider, if applicable
-        provider_options: Additional provider-specific configuration options
         structured_output_mode: The mode we should use to call the model for structured output, if it was trained with structured output.
         parser: A parser to use for the model, if applicable
         reasoning_capable: Whether the model is designed to output thinking in a structured format (eg <think></think>). If so we don't use COT across 2 calls, and ask for thinking and final response in the same call.
     """
 
     name: ModelProviderName
+    model_id: str | None = None
     supports_structured_output: bool = True
     supports_data_gen: bool = True
     untested_model: bool = False
     provider_finetune_id: str | None = None
-    provider_options: Dict = {}
     structured_output_mode: StructuredOutputMode = StructuredOutputMode.default
     parser: ModelParserID | None = None
     reasoning_capable: bool = False
+    supports_logprobs: bool = False
+
+    # TODO P1: 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
+    require_openrouter_reasoning: bool = False
+    logprobs_openrouter_options: bool = False
+    openrouter_skip_required_parameters: bool = False
+    thinking_level: Literal["low", "medium", "high"] | None = None
+    ollama_model_aliases: List[str] | None = None
+    anthropic_extended_thinking: bool = False
 
 
 class KilnModel(BaseModel):
@@ -152,14 +189,21 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openai,
-                provider_options={"model": "gpt-4o-mini"},
+                model_id="gpt-4o-mini",
                 provider_finetune_id="gpt-4o-mini-2024-07-18",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                supports_logprobs=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "openai/gpt-4o-mini"},
+                model_id="openai/gpt-4o-mini",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                supports_logprobs=True,
+                logprobs_openrouter_options=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="gpt-4o-mini",
             ),
         ],
     ),
@@ -171,14 +215,141 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openai,
-                provider_options={"model": "gpt-4o"},
+                model_id="gpt-4o",
                 provider_finetune_id="gpt-4o-2024-08-06",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                supports_logprobs=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "openai/gpt-4o"},
+                model_id="openai/gpt-4o",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_logprobs=True,
+                logprobs_openrouter_options=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="gpt-4o",
+            ),
+        ],
+    ),
+    # GPT o3 Mini Low
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o3_mini_low,
+        friendly_name="GPT o3 Mini - Low",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o3-mini",
+                thinking_level="low",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o3-mini",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="low",
+            ),
+        ],
+    ),
+    # GPT o3 Mini Medium
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o3_mini_medium,
+        friendly_name="GPT o3 Mini - Medium",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o3-mini",
+                thinking_level="medium",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o3-mini",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="medium",
+            ),
+        ],
+    ),
+    # GPT o3 Mini High
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o3_mini_high,
+        friendly_name="GPT o3 Mini - High",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o3-mini",
+                thinking_level="high",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o3-mini",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="high",
+            ),
+        ],
+    ),
+    # GPT o1 Low
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o1_low,
+        friendly_name="GPT o1 - Low",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o1",
+                thinking_level="low",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o1",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="low",
+            ),
+        ],
+    ),
+    # GPT o1 Medium
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o1_medium,
+        friendly_name="GPT o1 - Medium",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o1",
+                thinking_level="medium",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o1",
                 structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="medium",
+            ),
+        ],
+    ),
+    # GPT o1 High
+    KilnModel(
+        family=ModelFamily.gpt,
+        name=ModelName.gpt_o1_high,
+        friendly_name="GPT o1 - High",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openai,
+                model_id="o1",
+                thinking_level="high",
+                structured_output_mode=StructuredOutputMode.json_schema,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.azure_openai,
+                model_id="o1",
+                structured_output_mode=StructuredOutputMode.json_schema,
+                thinking_level="high",
             ),
         ],
     ),
@@ -191,7 +362,17 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={"model": "anthropic/claude-3-5-haiku"},
+                model_id="anthropic/claude-3-5-haiku",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.anthropic,
+                model_id="claude-3-5-haiku-20241022",
+                structured_output_mode=StructuredOutputMode.function_calling,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="claude-3-5-haiku",
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
             ),
         ],
     ),
@@ -204,50 +385,58 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={"model": "anthropic/claude-3.5-sonnet"},
+                model_id="anthropic/claude-3.5-sonnet",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.anthropic,
+                model_id="claude-3-5-sonnet-20241022",
+                structured_output_mode=StructuredOutputMode.function_calling,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="claude-3-5-sonnet",
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
             ),
         ],
     ),
-    # DeepSeek 3
+    # Claude 3.7 Sonnet
     KilnModel(
-        family=ModelFamily.deepseek,
-        name=ModelName.deepseek_3,
-        friendly_name="DeepSeek v3",
+        family=ModelFamily.claude,
+        name=ModelName.claude_3_7_sonnet,
+        friendly_name="Claude 3.7 Sonnet",
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "deepseek/deepseek-chat"},
+                structured_output_mode=StructuredOutputMode.function_calling,
+                model_id="anthropic/claude-3.7-sonnet",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.anthropic,
+                model_id="claude-3-7-sonnet-20250219",
                 structured_output_mode=StructuredOutputMode.function_calling,
             ),
         ],
     ),
-    # DeepSeek R1
+    # Claude 3.7 Sonnet Thinking
     KilnModel(
-        family=ModelFamily.deepseek,
-        name=ModelName.deepseek_r1,
-        friendly_name="DeepSeek R1",
+        family=ModelFamily.claude,
+        name=ModelName.claude_3_7_sonnet_thinking,
+        friendly_name="Claude 3.7 Sonnet Thinking",
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "deepseek/deepseek-r1"},
-                # No custom parser -- openrouter implemented it themselves
-                structured_output_mode=StructuredOutputMode.json_instructions,
+                model_id="anthropic/claude-3.7-sonnet:thinking",
                 reasoning_capable=True,
-            ),
-            KilnModelProvider(
-                name=ModelProviderName.fireworks_ai,
-                provider_options={"model": "accounts/fireworks/models/deepseek-r1"},
-                parser=ModelParserID.r1_thinking,
+                # For reasoning models, we need to use json_instructions with OpenRouter
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                reasoning_capable=True,
+                require_openrouter_reasoning=True,
             ),
             KilnModelProvider(
-                # I want your RAM
-                name=ModelProviderName.ollama,
-                provider_options={"model": "deepseek-r1:671b"},
-                parser=ModelParserID.r1_thinking,
-                structured_output_mode=StructuredOutputMode.json_instructions,
+                name=ModelProviderName.anthropic,
                 reasoning_capable=True,
+                model_id="claude-3-7-sonnet-20250219",
+                anthropic_extended_thinking=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
             ),
         ],
     ),
@@ -259,8 +448,18 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "google/gemini-pro-1.5"},
-                structured_output_mode=StructuredOutputMode.json_schema,
+                model_id="google/gemini-pro-1.5",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemini-1.5-pro",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="gemini-1.5-pro",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
         ],
     ),
@@ -272,8 +471,18 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "google/gemini-flash-1.5"},
-                structured_output_mode=StructuredOutputMode.json_schema,
+                model_id="google/gemini-flash-1.5",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemini-1.5-flash",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="gemini-1.5-flash",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
         ],
     ),
@@ -285,8 +494,15 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "google/gemini-flash-1.5-8b"},
-                structured_output_mode=StructuredOutputMode.json_mode,
+                model_id="google/gemini-flash-1.5-8b",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_data_gen=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemini-1.5-flash-8b",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_data_gen=False,
             ),
         ],
     ),
@@ -298,8 +514,18 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "google/gemini-2.0-flash-001"},
+                model_id="google/gemini-2.0-flash-001",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.gemini_api,
+                model_id="gemini-2.0-flash",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="gemini-2.0-flash",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
         ],
     ),
@@ -313,7 +539,7 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.openrouter,
                 supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={"model": "nvidia/llama-3.1-nemotron-70b-instruct"},
+                model_id="nvidia/llama-3.1-nemotron-70b-instruct",
             ),
         ],
     ),
@@ -325,39 +551,40 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.groq,
-                provider_options={"model": "llama-3.1-8b-instant"},
+                model_id="llama-3.1-8b-instant",
             ),
             KilnModelProvider(
                 name=ModelProviderName.amazon_bedrock,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                supports_data_gen=False,
-                provider_options={
-                    "model": "meta.llama3-1-8b-instruct-v1:0",
-                    "region_name": "us-west-2",  # Llama 3.1 only in west-2
-                },
+                supports_structured_output=False,
+                model_id="meta.llama3-1-8b-instruct-v1:0",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={
-                    "model": "llama3.1:8b",
-                    "model_aliases": ["llama3.1"],  # 8b is default
-                },
+                model_id="llama3.1:8b",
+                ollama_model_aliases=["llama3.1"],  # 8b is default
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_data_gen=False,
                 structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={"model": "meta-llama/llama-3.1-8b-instruct"},
+                model_id="meta-llama/llama-3.1-8b-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # JSON mode not ideal (no schema), but tool calling doesn't work on 8b
-                structured_output_mode=StructuredOutputMode.json_mode,
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_data_gen=False,
                 provider_finetune_id="accounts/fireworks/models/llama-v3p1-8b-instruct",
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p1-8b-instruct"
-                },
+                model_id="accounts/fireworks/models/llama-v3p1-8b-instruct",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
+                supports_data_gen=False,
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+                provider_finetune_id="meta-llama/Meta-Llama-3.1-8B-Instruct",
             ),
         ],
     ),
@@ -371,30 +598,35 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.amazon_bedrock,
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "meta.llama3-1-70b-instruct-v1:0",
-                    "region_name": "us-west-2",  # Llama 3.1 only in west-2
-                },
+                model_id="meta.llama3-1-70b-instruct-v1:0",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_data_gen=False,
-                structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={"model": "meta-llama/llama-3.1-70b-instruct"},
+                # Need to not pass "strict=True" to the function call to get this to work with logprobs for some reason. Openrouter issue.
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+                model_id="meta-llama/llama-3.1-70b-instruct",
+                supports_logprobs=True,
+                logprobs_openrouter_options=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "llama3.1:70b"},
+                model_id="llama3.1:70b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # Tool calling forces schema -- fireworks doesn't support json_schema, just json_mode
-                structured_output_mode=StructuredOutputMode.function_calling,
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
                 provider_finetune_id="accounts/fireworks/models/llama-v3p1-70b-instruct",
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p1-70b-instruct"
-                },
+                model_id="accounts/fireworks/models/llama-v3p1-70b-instruct",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo",
+                supports_data_gen=False,
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+                provider_finetune_id="meta-llama/Meta-Llama-3.1-70B-Instruct",
             ),
         ],
     ),
@@ -408,28 +640,29 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.amazon_bedrock,
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "meta.llama3-1-405b-instruct-v1:0",
-                    "region_name": "us-west-2",  # Llama 3.1 only in west-2
-                },
+                model_id="meta.llama3-1-405b-instruct-v1:0",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "llama3.1:405b"},
+                model_id="llama3.1:405b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={"model": "meta-llama/llama-3.1-405b-instruct"},
+                model_id="meta-llama/llama-3.1-405b-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p1-405b-instruct"
-                },
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+                model_id="accounts/fireworks/models/llama-v3p1-405b-instruct",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo",
+                supports_data_gen=False,
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
             ),
         ],
     ),
@@ -441,7 +674,7 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "mistralai/mistral-nemo"},
+                model_id="mistralai/mistral-nemo",
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
         ],
@@ -454,21 +687,18 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.amazon_bedrock,
-                structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={
-                    "model": "mistral.mistral-large-2407-v1:0",
-                    "region_name": "us-west-2",  # only in west-2
-                },
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                model_id="mistral.mistral-large-2407-v1:0",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "mistralai/mistral-large"},
+                model_id="mistralai/mistral-large",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "mistral-large"},
+                model_id="mistral-large",
             ),
         ],
     ),
@@ -480,7 +710,8 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.groq,
-                provider_options={"model": "llama-3.2-1b-preview"},
+                model_id="llama-3.2-1b-preview",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
                 supports_data_gen=False,
             ),
             KilnModelProvider(
@@ -488,13 +719,23 @@ built_in_models: List[KilnModel] = [
                 supports_structured_output=False,
                 supports_data_gen=False,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                provider_options={"model": "meta-llama/llama-3.2-1b-instruct"},
+                model_id="meta-llama/llama-3.2-1b-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={"model": "llama3.2:1b"},
+                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,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                provider_finetune_id="meta-llama/Llama-3.2-1B-Instruct",
             ),
         ],
     ),
@@ -506,7 +747,7 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.groq,
-                provider_options={"model": "llama-3.2-3b-preview"},
+                model_id="llama-3.2-3b-preview",
                 supports_data_gen=False,
             ),
             KilnModelProvider(
@@ -514,20 +755,32 @@ built_in_models: List[KilnModel] = [
                 supports_structured_output=False,
                 supports_data_gen=False,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "meta-llama/llama-3.2-3b-instruct"},
+                model_id="meta-llama/llama-3.2-3b-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 supports_data_gen=False,
-                provider_options={"model": "llama3.2"},
+                model_id="llama3.2",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 provider_finetune_id="accounts/fireworks/models/llama-v3p2-3b-instruct",
-                structured_output_mode=StructuredOutputMode.json_mode,
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p2-3b-instruct"
-                },
+                supports_structured_output=False,
+                supports_data_gen=False,
+                model_id="accounts/fireworks/models/llama-v3p2-3b-instruct",
+            ),
+            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",
+                supports_structured_output=False,
+                supports_data_gen=False,
+                provider_finetune_id="meta-llama/Llama-3.2-3B-Instruct",
             ),
         ],
     ),
@@ -539,25 +792,39 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.groq,
-                provider_options={"model": "llama-3.2-11b-vision-preview"},
+                model_id="llama-3.2-11b-vision-preview",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "meta-llama/llama-3.2-11b-vision-instruct"},
+                # Best mode, but fails to often to enable without warning
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_structured_output=False,
+                supports_data_gen=False,
+                model_id="meta-llama/llama-3.2-11b-vision-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "llama3.2-vision"},
+                model_id="llama3.2-vision",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p2-11b-vision-instruct"
-                },
-                structured_output_mode=StructuredOutputMode.json_mode,
+                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",
+                supports_structured_output=False,
+                supports_data_gen=False,
             ),
         ],
     ),
@@ -569,25 +836,30 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.groq,
-                provider_options={"model": "llama-3.2-90b-vision-preview"},
+                model_id="llama-3.2-90b-vision-preview",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "meta-llama/llama-3.2-90b-vision-instruct"},
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                model_id="meta-llama/llama-3.2-90b-vision-instruct",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "llama3.2-vision:90b"},
+                model_id="llama3.2-vision:90b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # No finetune support. https://docs.fireworks.ai/fine-tuning/fine-tuning-models
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p2-90b-vision-instruct"
-                },
-                structured_output_mode=StructuredOutputMode.json_mode,
+                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",
+                supports_structured_output=False,
+                supports_data_gen=False,
             ),
         ],
     ),
@@ -599,7 +871,7 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "meta-llama/llama-3.3-70b-instruct"},
+                model_id="meta-llama/llama-3.3-70b-instruct",
                 structured_output_mode=StructuredOutputMode.json_schema,
                 # Openrouter not working with json_schema or tools. JSON_schema sometimes works so force that, but not consistently so still not recommended.
                 supports_structured_output=False,
@@ -609,22 +881,32 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.groq,
                 supports_structured_output=True,
                 supports_data_gen=True,
-                provider_options={"model": "llama-3.3-70b-versatile"},
+                model_id="llama-3.3-70b-versatile",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "llama3.3"},
+                model_id="llama3.3",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
                 # Finetuning not live yet
                 # provider_finetune_id="accounts/fireworks/models/llama-v3p3-70b-instruct",
                 # Tool calling forces schema -- fireworks doesn't support json_schema, just json_mode
-                structured_output_mode=StructuredOutputMode.function_calling,
-                provider_options={
-                    "model": "accounts/fireworks/models/llama-v3p3-70b-instruct"
-                },
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+                model_id="accounts/fireworks/models/llama-v3p3-70b-instruct",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.vertex,
+                model_id="meta/llama-3.3-70b-instruct-maas",
+                # Doesn't work, TODO to debug
+                supports_structured_output=False,
+                supports_data_gen=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="meta-llama/Llama-3.3-70B-Instruct-Turbo",
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
             ),
         ],
     ),
@@ -639,23 +921,21 @@ built_in_models: List[KilnModel] = [
                 structured_output_mode=StructuredOutputMode.json_schema,
                 supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={"model": "phi3.5"},
+                model_id="phi3.5",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={"model": "microsoft/phi-3.5-mini-128k-instruct"},
+                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
-                structured_output_mode=StructuredOutputMode.json_mode,
+                supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "accounts/fireworks/models/phi-3-vision-128k-instruct"
-                },
+                model_id="accounts/fireworks/models/phi-3-vision-128k-instruct",
             ),
         ],
     ),
@@ -663,19 +943,45 @@ built_in_models: List[KilnModel] = [
     KilnModel(
         family=ModelFamily.phi,
         name=ModelName.phi_4,
-        friendly_name="Phi 4",
+        friendly_name="Phi 4 - 14B",
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 structured_output_mode=StructuredOutputMode.json_schema,
-                provider_options={"model": "phi4"},
+                model_id="phi4",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 # JSON mode not consistent enough to enable in UI
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
                 supports_data_gen=False,
-                provider_options={"model": "microsoft/phi-4"},
+                model_id="microsoft/phi-4",
+            ),
+        ],
+    ),
+    # Phi 4 5.6B
+    KilnModel(
+        family=ModelFamily.phi,
+        name=ModelName.phi_4_5p6b,
+        friendly_name="Phi 4 - 5.6B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="microsoft/phi-4-multimodal-instruct",
+                supports_structured_output=False,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Phi 4 Mini
+    KilnModel(
+        family=ModelFamily.phi,
+        name=ModelName.phi_4_mini,
+        friendly_name="Phi 4 Mini - 3.8B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="phi4-mini",
             ),
         ],
     ),
@@ -688,9 +994,7 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "gemma2:2b",
-                },
+                model_id="gemma2:2b",
             ),
         ],
     ),
@@ -703,15 +1007,15 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "gemma2:9b",
-                },
+                model_id="gemma2:9b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
+                # Best mode, but fails to often to enable without warning
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_structured_output=False,
                 supports_data_gen=False,
-                provider_options={"model": "google/gemma-2-9b-it"},
+                model_id="google/gemma-2-9b-it",
             ),
             # fireworks AI errors - not allowing system role. Exclude until resolved.
         ],
@@ -725,15 +1029,93 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 supports_data_gen=False,
-                provider_options={
-                    "model": "gemma2:27b",
-                },
+                model_id="gemma2:27b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
                 supports_data_gen=False,
-                provider_options={"model": "google/gemma-2-27b-it"},
+                model_id="google/gemma-2-27b-it",
+            ),
+        ],
+    ),
+    # Gemma 3 1B
+    KilnModel(
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3_1b,
+        friendly_name="Gemma 3 1B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3:1b",
+                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
+    KilnModel(
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3_4b,
+        friendly_name="Gemma 3 4B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3:4b",
+                ollama_model_aliases=["gemma3"],
+            ),
+            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",
+            ),
+        ],
+    ),
+    # Gemma 3 12B
+    KilnModel(
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3_12b,
+        friendly_name="Gemma 3 12B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3:12b",
+            ),
+            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",
+            ),
+        ],
+    ),
+    # Gemma 3 27B
+    KilnModel(
+        family=ModelFamily.gemma,
+        name=ModelName.gemma_3_27b,
+        friendly_name="Gemma 3 27B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="gemma3:27b",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                model_id="google/gemma-3-27b-it",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.huggingface,
+                model_id="google/gemma-3-27b-it",
+                structured_output_mode=StructuredOutputMode.json_instructions,
             ),
         ],
     ),
@@ -745,13 +1127,58 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "mistralai/mixtral-8x7b-instruct"},
+                model_id="mistralai/mixtral-8x7b-instruct",
                 supports_data_gen=False,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
-                provider_options={"model": "mixtral"},
+                model_id="mixtral",
+            ),
+        ],
+    ),
+    # QwQ 32B
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwq_32b,
+        friendly_name="QwQ 32B (Qwen Reasoning)",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="qwen/qwq-32b",
+                reasoning_capable=True,
+                require_openrouter_reasoning=True,
+                r1_openrouter_options=True,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/qwq-32b",
+                reasoning_capable=True,
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwq",
+                reasoning_capable=True,
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.groq,
+                model_id="qwen-qwq-32b",
+                reasoning_capable=True,
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="Qwen/QwQ-32B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+                reasoning_capable=True,
             ),
         ],
     ),
@@ -763,12 +1190,29 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "qwen/qwen-2.5-7b-instruct"},
+                model_id="qwen/qwen-2.5-7b-instruct",
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
-                provider_options={"model": "qwen2.5"},
+                model_id="qwen2.5",
+            ),
+        ],
+    ),
+    # Qwen 2.5 14B
+    KilnModel(
+        family=ModelFamily.qwen,
+        name=ModelName.qwen_2p5_14b,
+        friendly_name="Qwen 2.5 14B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                provider_finetune_id="Qwen/Qwen2.5-14B-Instruct",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                model_id="qwen2.5:14b",
+                supports_data_gen=False,
             ),
         ],
     ),
@@ -780,7 +1224,7 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
-                provider_options={"model": "qwen/qwen-2.5-72b-instruct"},
+                model_id="qwen/qwen-2.5-72b-instruct",
                 # Not consistent with structure data. Works sometimes but not often
                 supports_structured_output=False,
                 supports_data_gen=False,
@@ -788,17 +1232,17 @@ built_in_models: List[KilnModel] = [
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
-                provider_options={"model": "qwen2.5:72b"},
+                model_id="qwen2.5:72b",
             ),
             KilnModelProvider(
                 name=ModelProviderName.fireworks_ai,
-                provider_options={
-                    "model": "accounts/fireworks/models/qwen2p5-72b-instruct"
-                },
-                # Fireworks will start tuning, but it never finishes.
-                # provider_finetune_id="accounts/fireworks/models/qwen2p5-72b-instruct",
+                model_id="accounts/fireworks/models/qwen2p5-72b-instruct",
                 # Tool calling forces schema -- fireworks doesn't support json_schema, just json_mode
-                structured_output_mode=StructuredOutputMode.function_calling,
+                structured_output_mode=StructuredOutputMode.function_calling_weak,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                provider_finetune_id="Qwen/Qwen2.5-72B-Instruct",
             ),
         ],
     ),
@@ -811,11 +1255,75 @@ built_in_models: List[KilnModel] = [
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
                 structured_output_mode=StructuredOutputMode.json_instruction_and_object,
-                provider_options={"model": "mistralai/mistral-small-24b-instruct-2501"},
+                model_id="mistralai/mistral-small-24b-instruct-2501",
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
-                provider_options={"model": "mistral-small:24b"},
+                model_id="mistral-small:24b",
+            ),
+        ],
+    ),
+    # DeepSeek 3
+    KilnModel(
+        family=ModelFamily.deepseek,
+        name=ModelName.deepseek_3,
+        friendly_name="DeepSeek V3",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="deepseek/deepseek-chat",
+                structured_output_mode=StructuredOutputMode.function_calling,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/deepseek-v3",
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                supports_structured_output=True,
+                supports_data_gen=False,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="deepseek-ai/DeepSeek-V3",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+            ),
+        ],
+    ),
+    # DeepSeek R1
+    KilnModel(
+        family=ModelFamily.deepseek,
+        name=ModelName.deepseek_r1,
+        friendly_name="DeepSeek R1",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="deepseek/deepseek-r1",
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.fireworks_ai,
+                model_id="accounts/fireworks/models/deepseek-r1",
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=True,
+            ),
+            KilnModelProvider(
+                # I want your RAM
+                name=ModelProviderName.ollama,
+                model_id="deepseek-r1:671b",
+                parser=ModelParserID.r1_thinking,
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                reasoning_capable=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,
             ),
         ],
     ),
@@ -829,14 +1337,24 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.openrouter,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek/deepseek-r1-distill-qwen-32b"},
+                model_id="deepseek/deepseek-r1-distill-qwen-32b",
+                r1_openrouter_options=True,
+                parser=ModelParserID.r1_thinking,
+                require_openrouter_reasoning=True,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:32b"},
+                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,
+                reasoning_capable=True,
             ),
         ],
     ),
@@ -850,7 +1368,10 @@ built_in_models: List[KilnModel] = [
                 name=ModelProviderName.openrouter,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek/deepseek-r1-distill-llama-70b"},
+                model_id="deepseek/deepseek-r1-distill-llama-70b",
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+                parser=ModelParserID.r1_thinking,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
@@ -858,7 +1379,13 @@ built_in_models: List[KilnModel] = [
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:70b"},
+                model_id="deepseek-r1:70b",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
             ),
         ],
     ),
@@ -873,7 +1400,11 @@ built_in_models: List[KilnModel] = [
                 supports_data_gen=False,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek/deepseek-r1-distill-qwen-14b"},
+                model_id="deepseek/deepseek-r1-distill-qwen-14b",
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+                openrouter_skip_required_parameters=True,
+                parser=ModelParserID.r1_thinking,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
@@ -881,7 +1412,13 @@ built_in_models: List[KilnModel] = [
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:14b"},
+                model_id="deepseek-r1:14b",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
             ),
         ],
     ),
@@ -893,18 +1430,26 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.openrouter,
+                supports_structured_output=False,
                 supports_data_gen=False,
                 reasoning_capable=True,
+                # Best mode, but fails to often to enable without warning
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek/deepseek-r1-distill-llama-8b"},
+                model_id="deepseek/deepseek-r1-distill-llama-8b",
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+                openrouter_skip_required_parameters=True,
+                parser=ModelParserID.r1_thinking,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
+                supports_structured_output=False,
                 supports_data_gen=False,
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
+                # Best mode, but fails to often to enable without warning
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:8b"},
+                model_id="deepseek-r1:8b",
             ),
         ],
     ),
@@ -916,11 +1461,13 @@ built_in_models: List[KilnModel] = [
         providers=[
             KilnModelProvider(
                 name=ModelProviderName.ollama,
+                # Best mode, but fails to often to enable without warning
+                supports_structured_output=False,
                 supports_data_gen=False,
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:7b"},
+                model_id="deepseek-r1:7b",
             ),
         ],
     ),
@@ -936,15 +1483,63 @@ built_in_models: List[KilnModel] = [
                 supports_data_gen=False,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek/deepseek-r1-distill-qwen-1.5b"},
+                model_id="deepseek/deepseek-r1-distill-qwen-1.5b",
+                r1_openrouter_options=True,
+                require_openrouter_reasoning=True,
+                openrouter_skip_required_parameters=True,
+                parser=ModelParserID.r1_thinking,
             ),
             KilnModelProvider(
                 name=ModelProviderName.ollama,
+                supports_structured_output=False,
                 supports_data_gen=False,
                 parser=ModelParserID.r1_thinking,
                 reasoning_capable=True,
                 structured_output_mode=StructuredOutputMode.json_instructions,
-                provider_options={"model": "deepseek-r1:1.5b"},
+                model_id="deepseek-r1:1.5b",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.together_ai,
+                model_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
+                structured_output_mode=StructuredOutputMode.json_instructions,
+                parser=ModelParserID.r1_thinking,
+                supports_structured_output=False,
+                supports_data_gen=False,
+            ),
+        ],
+    ),
+    # Dolphin 2.9 Mixtral 8x22B
+    KilnModel(
+        family=ModelFamily.dolphin,
+        name=ModelName.dolphin_2_9_8x22b,
+        friendly_name="Dolphin 2.9 8x22B",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.ollama,
+                structured_output_mode=StructuredOutputMode.json_schema,
+                supports_data_gen=True,
+                model_id="dolphin-mixtral:8x22b",
+            ),
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_instruction_and_object,
+                model_id="cognitivecomputations/dolphin-mixtral-8x22b",
+            ),
+        ],
+    ),
+    # Grok 2
+    KilnModel(
+        family=ModelFamily.grok,
+        name=ModelName.grok_2,
+        friendly_name="Grok 2",
+        providers=[
+            KilnModelProvider(
+                name=ModelProviderName.openrouter,
+                model_id="x-ai/grok-2-1212",
+                supports_structured_output=True,
+                supports_data_gen=True,
+                structured_output_mode=StructuredOutputMode.json_schema,
             ),
         ],
     ),