@bendyline/gezel-inference 1.0.1 → 1.0.3

package/dist/catalog/chat-models/ne/nemotron3.5-bolt-74b-iq2/manifest.json

@@ -0,0 +1,157 @@
+{
+  "schemaVersion": 1,
+  "kind": "chat-model",
+  "id": "nemotron3.5-bolt-74b-iq2",
+  "name": "Nemotron 3.5 Bolt (74B-A9B, 2-bit)",
+  "description": "NVIDIA's Nemotron 3.5 Bolt: Super VL distilled down to 74B parameters with 9B active, quantization-aware trained to fit in about 24GB. Native reasoning, tool calling, image understanding, and a 256K context window, with a small LiLiCorr drafter that speeds up replies. Needs an NVIDIA or Vulkan GPU with about 26GB of memory; not offered on Macs.",
+  "tags": [
+    "nvidia",
+    "reasoning",
+    "agentic",
+    "coding",
+    "tools",
+    "vision",
+    "long-context",
+    "moe",
+    "hybrid",
+    "mamba",
+    "open-weights"
+  ],
+  "category": "reasoning",
+  "maintainer": {
+    "name": "NVIDIA",
+    "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB"
+  },
+  "version": "1.0.0",
+  "updatedAt": "2026-10-17T00:00:00Z",
+  "license": "OpenMDW-1.1",
+  "licenseClass": "open",
+  "licenseShortName": "OpenMDW 1.1",
+  "licenseUrl": "https://openmdw.ai/license/1-1/",
+  "parameterSize": "74B",
+  "approxSizeBytes": 22018326528,
+  "supportsTools": true,
+  "contextWindow": 262144,
+  "kvBytesPerTokenF16": 8192,
+  "upstream": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
+  "style": {
+    "family": "nemotron",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.private-reasoning-guidance",
+    "prompt.tool-cookbook-condensed",
+    "mcp.compact-tool-schemas",
+    "fabrication.detect-past-tense-no-tools",
+    "prompt.retrieval-first",
+    "prompt.workspace-gestalt",
+    "turn.preamble-folding",
+    "turn.ramble-detection"
+  ],
+  "tuning": {
+    "sampling": {
+      "temperature": 1,
+      "topP": 0.95,
+      "maxTokens": 16384
+    },
+    "samplingWhenThinking": {
+      "temperature": 1,
+      "topP": 0.95
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 8192
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 8192
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 32768
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 12288
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.9,
+          "seed": 0,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 6144
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 0.95,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    },
+    "engine": {
+      "llamaCpp": {
+        "spec": {
+          "type": "draft-dflash",
+          "nMax": 7
+        }
+      }
+    }
+  },
+  "minGezelVersion": "1.26290",
+  "llamaCpp": {
+    "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF",
+    "revision": "e0fecb69cb3e5e4c5e8a056fa2aec98c78e91246",
+    "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+    "sha256": "91d63363e782c8a3d5f4be42816eba8e0d42dd2530bd07a7d0d2dc440599af75",
+    "approxSizeBytes": 22018326528,
+    "quantization": "IQ2_XXS",
+    "mmproj": {
+      "filename": "mmproj-NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+      "sha256": "0ee891eadc7bf0cae6289b249c74eb1ef420534351f6ef8b64d49f0ca6a926b0",
+      "sizeBytes": 1733692736
+    },
+    "draftModel": {
+      "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF",
+      "revision": "182d2987939e12f458298b5d5f6ef98912f4ad6c",
+      "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr.gguf",
+      "sha256": "080e1ea41c9562a7804dbf30c6fbd4d17e216c6b8ff666f88e6edb1816010194",
+      "sizeBytes": 998242560
+    },
+    "excludeBackends": [
+      "metal"
+    ],
+    "excludeReason": "llama.cpp's Metal backend has no NVFP4 kernels and runs these layers on the CPU, so this model would be far too slow on a Mac."
+  }
+}

package/dist/catalog/chat-models/ne/nemotron3.5-bolt-74b-iq2/versions/1.0.0/manifest.json

@@ -0,0 +1,30 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-17T00:00:00Z",
+  "approxSizeBytes": 22018326528,
+  "llamaCpp": {
+    "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF",
+    "revision": "e0fecb69cb3e5e4c5e8a056fa2aec98c78e91246",
+    "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+    "sha256": "91d63363e782c8a3d5f4be42816eba8e0d42dd2530bd07a7d0d2dc440599af75",
+    "approxSizeBytes": 22018326528,
+    "quantization": "IQ2_XXS",
+    "mmproj": {
+      "filename": "mmproj-NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+      "sha256": "0ee891eadc7bf0cae6289b249c74eb1ef420534351f6ef8b64d49f0ca6a926b0",
+      "sizeBytes": 1733692736
+    },
+    "draftModel": {
+      "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF",
+      "revision": "182d2987939e12f458298b5d5f6ef98912f4ad6c",
+      "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr.gguf",
+      "sha256": "080e1ea41c9562a7804dbf30c6fbd4d17e216c6b8ff666f88e6edb1816010194",
+      "sizeBytes": 998242560
+    },
+    "excludeBackends": [
+      "metal"
+    ],
+    "excludeReason": "llama.cpp's Metal backend has no NVFP4 kernels and runs these layers on the CPU, so this model would be far too slow on a Mac."
+  }
+}

package/dist/catalog/chat-models/ne/nemotron3.5-super-vl-120b-nvfp4/manifest.json

@@ -0,0 +1,156 @@
+{
+  "schemaVersion": 1,
+  "kind": "chat-model",
+  "id": "nemotron3.5-super-vl-120b-nvfp4",
+  "name": "Nemotron 3.5 Super VL (120B-A12B, NVFP4)",
+  "description": "NVIDIA's flagship Nemotron 3.5 Super VL at 120B parameters with 12B active, in NVIDIA's quantization-aware NVFP4. Native reasoning, tool calling, image understanding, and a 256K context window, with its own multi-token-prediction head for faster replies. Workstation-tier: needs about 90GB of usable memory, and runs best on NVIDIA Blackwell GPUs such as DGX Spark; not offered on Macs.",
+  "tags": [
+    "nvidia",
+    "reasoning",
+    "agentic",
+    "coding",
+    "tools",
+    "vision",
+    "long-context",
+    "moe",
+    "hybrid",
+    "mamba",
+    "open-weights",
+    "large"
+  ],
+  "category": "reasoning",
+  "maintainer": {
+    "name": "NVIDIA",
+    "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4"
+  },
+  "version": "1.0.0",
+  "updatedAt": "2026-10-17T00:00:00Z",
+  "license": "OpenMDW-1.1",
+  "licenseClass": "open",
+  "licenseShortName": "OpenMDW 1.1",
+  "licenseUrl": "https://openmdw.ai/license/1-1/",
+  "parameterSize": "120B",
+  "approxSizeBytes": 76918795552,
+  "supportsTools": true,
+  "contextWindow": 262144,
+  "kvBytesPerTokenF16": 8192,
+  "upstream": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4",
+  "style": {
+    "family": "nemotron",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.private-reasoning-guidance",
+    "prompt.tool-cookbook-condensed",
+    "mcp.compact-tool-schemas",
+    "fabrication.detect-past-tense-no-tools",
+    "prompt.retrieval-first",
+    "prompt.workspace-gestalt",
+    "turn.preamble-folding",
+    "turn.ramble-detection"
+  ],
+  "tuning": {
+    "sampling": {
+      "temperature": 1,
+      "topP": 0.95,
+      "maxTokens": 16384
+    },
+    "samplingWhenThinking": {
+      "temperature": 1,
+      "topP": 0.95
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 8192
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 8192
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 32768
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 12288
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.9,
+          "seed": 0,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 6144
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 0.95,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    },
+    "engine": {
+      "llamaCpp": {
+        "spec": {
+          "mtp": true,
+          "nMax": 4
+        }
+      }
+    }
+  },
+  "minGezelVersion": "1.26290",
+  "llamaCpp": {
+    "huggingfaceRepo": "ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
+    "revision": "3c91ad1be60865386af92e8147dafb32470b097d",
+    "filename": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4.gguf",
+    "sha256": "7611c78aabe730c46ddf805fb6f3f2d2fe430a53495451c319b561377137dedd",
+    "approxSizeBytes": 76918795552,
+    "quantization": "NVFP4",
+    "mmproj": {
+      "filename": "mmproj-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+      "sha256": "5687736375f1056b25441f07293ad043107fe8418ed62795da8cae1e61f0aa88",
+      "sizeBytes": 966925568
+    },
+    "draftModel": {
+      "filename": "mtp-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+      "sha256": "fe4564dd41d934446d8168c4d14a61c262e7b19618f953318ee049f59621f9c6",
+      "sizeBytes": 4276991712
+    },
+    "excludeBackends": [
+      "metal"
+    ],
+    "excludeReason": "llama.cpp's Metal backend has no NVFP4 kernels and runs these layers on the CPU, so this model would be far too slow on a Mac."
+  }
+}

package/dist/catalog/chat-models/ne/nemotron3.5-super-vl-120b-nvfp4/versions/1.0.0/manifest.json

@@ -0,0 +1,28 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-17T00:00:00Z",
+  "approxSizeBytes": 76918795552,
+  "llamaCpp": {
+    "huggingfaceRepo": "ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
+    "revision": "3c91ad1be60865386af92e8147dafb32470b097d",
+    "filename": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4.gguf",
+    "sha256": "7611c78aabe730c46ddf805fb6f3f2d2fe430a53495451c319b561377137dedd",
+    "approxSizeBytes": 76918795552,
+    "quantization": "NVFP4",
+    "mmproj": {
+      "filename": "mmproj-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+      "sha256": "5687736375f1056b25441f07293ad043107fe8418ed62795da8cae1e61f0aa88",
+      "sizeBytes": 966925568
+    },
+    "draftModel": {
+      "filename": "mtp-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+      "sha256": "fe4564dd41d934446d8168c4d14a61c262e7b19618f953318ee049f59621f9c6",
+      "sizeBytes": 4276991712
+    },
+    "excludeBackends": [
+      "metal"
+    ],
+    "excludeReason": "llama.cpp's Metal backend has no NVFP4 kernels and runs these layers on the CPU, so this model would be far too slow on a Mac."
+  }
+}