@bendyline/gilde 0.1.43 → 0.1.45

package/authoring/chat-models/deepseek-r1-8b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "deepseek-r1-8b-q4",
-  "name": "DeepSeek R1 (8B)",
+  "name": "DeepSeek R1 (8B, Q4)",
   "description": "Reasoning-tuned 8B model. Slower than a plain chat model, but thinks through its answers before replying.",
   "tags": [
     "deepseek",

package/authoring/chat-models/gemma4-31b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "gemma4-31b-q4",
-  "name": "Gemma 4 (31B)",
+  "name": "Gemma 4 (31B, Q4)",
   "description": "Google's flagship Gemma 4 at 30.7B parameters. Multimodal (text + image + video), 256K context, configurable thinking mode, native function-calling, 140+ languages. Frontier-tier instruction-following on device — needs 32GB+ of usable memory.",
   "tags": [
     "google",

package/authoring/chat-models/gpt-oss-120b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "gpt-oss-120b-q4",
-  "name": "GPT-OSS (120B, MoE)",
+  "name": "GPT-OSS (120B, MoE, Q4)",
   "description": "OpenAI's open-weight 117B-parameter Mixture-of-Experts model (~5.1B active per token). Configurable reasoning effort (low/medium/high), full chain-of-thought access, and native function-calling. Frontier-class on-device reasoning for workstations and Mac Studio-class hardware — needs ~80GB of usable memory.",
   "tags": [
     "openai",

package/authoring/chat-models/gpt-oss-20b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "gpt-oss-20b-q4",
-  "name": "GPT-OSS (20B, MoE)",
+  "name": "GPT-OSS (20B, MoE, Q4)",
   "description": "OpenAI's open-weight 20B-parameter Mixture-of-Experts model (~3.6B active per token). Configurable reasoning effort (low/medium/high), full chain-of-thought access, and native function-calling. Lower-latency frontier-class on-device option for machines with 16GB+ memory.",
   "tags": [
     "openai",

package/authoring/chat-models/lfm2.5-2.6b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "lfm2.5-2.6b-q4",
-  "name": "LFM2.5 (2.6B)",
+  "name": "LFM2.5 (2.6B, Q4)",
   "description": "Liquid AI's LFM2.5 at 2.6B parameters. A hybrid convolution-attention design built for on-device speed — light on memory, quick even on modest hardware, with integrated reasoning, native tool calling, and a 128K context window.",
   "tags": [
     "liquid-ai",

package/authoring/chat-models/llama3.2-3b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "llama3.2-3b-q4",
-  "name": "Llama 3.2 (3B)",
+  "name": "Llama 3.2 (3B, Q4)",
   "description": "Meta's 3B chat model. Small enough to run on modest hardware, sharp enough to be a useful default.",
   "tags": [
     "meta",

package/authoring/chat-models/mistral-7b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "mistral-7b-q4",
-  "name": "Mistral 7B Instruct",
+  "name": "Mistral 7B Instruct (Q4)",
   "description": "Well-rounded 7B instruct model from Mistral AI. Good general-purpose default.",
   "tags": [
     "mistral",

package/authoring/chat-models/mistral-medium-3.5-128b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "mistral-medium-3.5-128b-q4",
-  "name": "Mistral Medium 3.5 (128B)",
+  "name": "Mistral Medium 3.5 (128B, Q4)",
   "description": "Mistral AI's frontier dense 128B model with configurable reasoning (instant or chain-of-thought via `reasoning_effort`). Native function-calling, 256K context. Workstation-tier on-device — needs ~80GB of usable memory for Q4_K_M; an MLX 4-bit build is available for Apple Silicon Mac Studio-class hardware. License is a Modified MIT with a revenue carve-out — free for commercial and non-commercial use unless the licensee company exceeds Mistral's revenue threshold.",
   "tags": [
     "mistral",

package/authoring/chat-models/nemotron3-nano-30b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "nemotron3-nano-30b-q4",
-  "name": "Nemotron-3 Nano (30B, MoE Reasoning)",
+  "name": "Nemotron-3 Nano (30B, MoE Reasoning, Q4)",
   "description": "NVIDIA's Nemotron-3 Nano Omni at 30B parameters, tuned for reasoning. Multimodal (text + image + audio + video), 128K context, native tool calling. Capable-tier on-device option for machines with 32GB+ memory.",
   "tags": [
     "nvidia",

package/authoring/chat-models/nemotron3-super-120b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "nemotron3-super-120b-q4",
-  "name": "Nemotron-3 Super (120B-A12B)",
+  "name": "Nemotron-3 Super (120B-A12B, Q4)",
   "description": "NVIDIA's flagship Nemotron-3 Super at 120B parameters, tuned for reasoning. Native tool calling, 128K context. Frontier-tier on-device reasoning for workstations and Mac Studio-class hardware — needs ~100GB of usable memory. No Apple Silicon build yet.",
   "tags": [
     "nvidia",

package/authoring/chat-models/ornith1.5-35b-a3b-q4.json

@@ -0,0 +1,150 @@
+{
+  "id": "ornith1.5-35b-a3b-q4",
+  "name": "Ornith 1.5 (35B-A3B, Q4)",
+  "description": "Ornith AI's Ornith 1.5 at 35B total parameters with roughly 3B active per token — a self-improving, agentic-coding model trained for software engineering, tool use, and reasoning. The 4-bit build is the practical pick at this size, keeping fast MoE decode in a much smaller footprint than Q8, with native tool calling and a 256K context window. MIT licensed; best suited to machines with 32GB+ usable memory.",
+  "tags": [
+    "ornith",
+    "coding",
+    "agentic",
+    "tools",
+    "reasoning",
+    "long-context",
+    "moe"
+  ],
+  "category": "coding",
+  "maintainer": {
+    "name": "Ornith AI",
+    "url": "https://huggingface.co/ornith-ai"
+  },
+  "version": "1.0.0",
+  "updatedAt": "2026-08-27T00:00:00Z",
+  "versionMinGezelVersion": "1.26227",
+  "license": "MIT",
+  "licenseClass": "open",
+  "licenseShortName": "MIT",
+  "parameterSize": "35B",
+  "approxSizeBytes": 21713463040,
+  "supportsTools": true,
+  "contextWindow": 262144,
+  "kvBytesPerTokenF16": 20480,
+  "kvFixedBytesF16": 64389120,
+  "upstream": "https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B",
+  "style": {
+    "family": "qwen",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.private-reasoning-guidance",
+    "prompt.tool-cookbook-condensed",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.ollama-num-predict-bumped",
+    "turn.preamble-folding",
+    "turn.ramble-detection",
+    "mcp.compact-tool-schemas",
+    "tools.mlx-grammar"
+  ],
+  "tuning": {
+    "sampling": {
+      "temperature": 0.6,
+      "topP": 0.95,
+      "topK": 20,
+      "minP": 0,
+      "repetitionPenalty": 1,
+      "maxTokens": 8192
+    },
+    "samplingWhenThinking": {
+      "temperature": 0.6,
+      "topP": 0.95
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 4096
+    },
+    "promptTags": {
+      "enableThinkingTag": "/think",
+      "disableThinkingTag": "/no_think"
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 12288
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 4096
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 6144
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.4,
+          "topP": 0.9,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "seed": 0,
+          "maxTokens": 6144
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 4096
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 0.7,
+          "topP": 0.8,
+          "topK": 20,
+          "minP": 0,
+          "presencePenalty": 1.5,
+          "repetitionPenalty": 1,
+          "maxTokens": 4096
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1.1,
+          "topP": 0.95,
+          "topK": 40,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "ornith-ai/Ornith-1.5-35B-A3B-GGUF",
+    "filename": "Ornith-1.5-35B-Q4_K_M.gguf",
+    "quantization": "Q4_K_M"
+  },
+  "mlx": {
+    "huggingfaceRepo": "ornith-ai/Ornith-1.5-35B-A3B-MLX-4bit",
+    "quantization": "4bit",
+    "disabledReason": "The official MLX conversion omits 393 vision-tower parameters while declaring the multimodal Qwen3.5 architecture, so mlx-vlm cannot construct the checkpoint."
+  }
+}

package/authoring/chat-models/qwen3.5-2b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "qwen3.5-2b-q4",
-  "name": "Qwen 3.5 (2B)",
+  "name": "Qwen 3.5 (2B, Q4)",
   "description": "Alibaba's Qwen 3.5 at 2B parameters. Multimodal (text + images) with native tool calling and a 256K context window. A very compact choice for constrained devices.",
   "tags": [
     "alibaba",

package/authoring/chat-models/qwen3.5-4b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "qwen3.5-4b-q4",
-  "name": "Qwen 3.5 (4B)",
+  "name": "Qwen 3.5 (4B, Q4)",
   "description": "Alibaba's Qwen 3.5 at 4B parameters. Multimodal (text + images) with native tool calling and a 256K context window. A balanced compact default for modern laptops.",
   "tags": [
     "alibaba",

package/authoring/chat-models/qwen3.5-9b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "qwen3.5-9b-q4",
-  "name": "Qwen 3.5 (9B)",
+  "name": "Qwen 3.5 (9B, Q4)",
   "description": "Alibaba's Qwen 3.5 at 9B parameters. Multimodal (text + images), native tool calling, 256K context. Strong reasoning + tool use for devices with a 12GB+ inference budget — the small-tier sweet spot for tool-heavy work.",
   "tags": [
     "alibaba",

package/authoring/chat-models/talkie-1930-13b-q4.json

@@ -1,6 +1,6 @@
 {
   "id": "talkie-1930-13b-q4",
-  "name": "Talkie 1930 (13B)",
+  "name": "Talkie 1930 (13B, Q4)",
   "description": "A 13B instruct model trained solely on pre-1931 English-language text — it writes and converses in an authentic period voice, free of modern anachronisms. Best for creative writing, correspondence, and in-character chat rather than tool use or current-events questions. Runs a community GGUF that rescales talkie's out-of-range layer-14 ffn_down channel so K-quantization stays coherent. 2K context. Apache-2.0.",
   "tags": [
     "talkie-lm",

package/data/chat-models/de/deepseek-r1-8b-q4/manifest.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "kind": "chat-model",
   "id": "deepseek-r1-8b-q4",
-  "name": "DeepSeek R1 (8B)",
+  "name": "DeepSeek R1 (8B, Q4)",
   "description": "Reasoning-tuned 8B model. Slower than a plain chat model, but thinks through its answers before replying.",
   "tags": [
     "deepseek",

package/data/chat-models/ge/gemma4-31b-q4/manifest.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "kind": "chat-model",
   "id": "gemma4-31b-q4",
-  "name": "Gemma 4 (31B)",
+  "name": "Gemma 4 (31B, Q4)",
   "description": "Google's flagship Gemma 4 at 30.7B parameters. Multimodal (text + image + video), 256K context, configurable thinking mode, native function-calling, 140+ languages. Frontier-tier instruction-following on device — needs 32GB+ of usable memory.",
   "tags": [
     "google",

package/data/chat-models/gp/gpt-oss-120b-q4/manifest.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "kind": "chat-model",
   "id": "gpt-oss-120b-q4",
-  "name": "GPT-OSS (120B, MoE)",
+  "name": "GPT-OSS (120B, MoE, Q4)",
   "description": "OpenAI's open-weight 117B-parameter Mixture-of-Experts model (~5.1B active per token). Configurable reasoning effort (low/medium/high), full chain-of-thought access, and native function-calling. Frontier-class on-device reasoning for workstations and Mac Studio-class hardware — needs ~80GB of usable memory.",
   "tags": [
     "openai",

package/data/chat-models/gp/gpt-oss-20b-q4/manifest.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "kind": "chat-model",
   "id": "gpt-oss-20b-q4",
-  "name": "GPT-OSS (20B, MoE)",
+  "name": "GPT-OSS (20B, MoE, Q4)",
   "description": "OpenAI's open-weight 20B-parameter Mixture-of-Experts model (~3.6B active per token). Configurable reasoning effort (low/medium/high), full chain-of-thought access, and native function-calling. Lower-latency frontier-class on-device option for machines with 16GB+ memory.",
   "tags": [
     "openai",