@bendyline/gilde 0.2.2 → 0.2.4

package/authoring/chat-models/btl4-35b-q4.json

@@ -35,6 +35,7 @@
     "toolCallFormat": "function-call"
   },
   "behaviors": [
+    "provider.merge-system-messages",
     "reasoning.strip-think-tags",
     "prompt.private-reasoning-guidance",
     "prompt.tool-cookbook-condensed",

package/authoring/chat-models/glm-5.3-flash-320b-q2.json

@@ -26,9 +26,9 @@
     "name": "Z.ai",
     "url": "https://huggingface.co/zai-org/GLM-5.3-Flash"
   },
-  "version": "1.0.1",
+  "version": "1.0.2",
   "versionMinGezelVersion": "1.26248",
-  "updatedAt": "2026-09-06T00:00:00Z",
+  "updatedAt": "2026-10-17T00:00:00Z",
   "license": "MIT",
   "licenseClass": "open",
   "licenseShortName": "MIT",
@@ -122,8 +122,9 @@
     "filename": "GLM-5.3-Flash-Q2.gguf",
     "visionEncoderFilename": "GLM-5.3-Flash-Vision-Encoder.gguf",
     "quantization": "IQ2_XXS/Q2_K",
-    "residentBytes": 100865554432,
-    "ssdStreaming": false,
+    "residentBytes": 47244640256,
+    "cacheExpertsBytes": 34359738368,
+    "ssdStreaming": true,
     "maxLaunchCtx": 32768,
     "mtp": {
       "exactSampling": true

package/authoring/chat-models/minicpm-v4.7-35b-a3b-q4.json

@@ -1,7 +1,7 @@
 {
   "id": "minicpm-v4.7-35b-a3b-q4",
   "name": "MiniCPM-V 4.7 (35B-A3B, Q4)",
-  "description": "OpenBMB's vision-language model with 35B total parameters and roughly 3B active per token. Understands images, documents, and multiple images, with native tool calling and optional thinking. The Q4_K_M download includes the 16x vision projector. Requires llama.cpp with MiniCPM-V 4.7 support (upstream PR #29416); Gezel's v0.6.0 engine cannot run its vision encoder.",
+  "description": "OpenBMB's vision-language model with 35B total parameters and roughly 3B active per token. Understands images, documents, and multiple images, with native tool calling and optional thinking. The Q4_K_M download includes the 16x vision projector. Requires llama.cpp with MiniCPM-V 4.7 support (upstream PR #29416), which Gezel's bundled engine does not have yet; release builds hide it until one does.",
   "tags": [
     "openbmb",
     "vision",
@@ -24,12 +24,14 @@
   "supportsTools": true,
   "contextWindow": 262144,
   "upstream": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B",
+  "versionMinGezelVersion": "1.99999",
   "style": {
     "family": "qwen",
     "reasoningFormat": "think",
     "toolCallFormat": "function-call"
   },
   "behaviors": [
+    "provider.merge-system-messages",
     "reasoning.strip-think-tags",
     "mcp.compact-tool-schemas",
     "prompt.tool-cookbook-condensed",

package/authoring/chat-models/nemotron3.5-bolt-74b-iq2.json

@@ -0,0 +1,145 @@
+{
+  "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",
+  "minGezelVersion": "1.26290",
+  "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
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF",
+    "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+    "quantization": "IQ2_XXS",
+    "mmprojFilename": "mmproj-NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf",
+    "draftModel": {
+      "huggingfaceRepo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF",
+      "filename": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr.gguf"
+    },
+    "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/authoring/chat-models/nemotron3.5-super-vl-120b-nvfp4.json

@@ -0,0 +1,143 @@
+{
+  "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",
+  "minGezelVersion": "1.26290",
+  "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
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
+    "filename": "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4.gguf",
+    "quantization": "NVFP4",
+    "mmprojFilename": "mmproj-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+    "draftModelFilename": "mtp-NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q8_0.gguf",
+    "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/data/chat-models/bt/btl4-35b-q4/manifest.json

@@ -37,6 +37,7 @@
     "toolCallFormat": "function-call"
   },
   "behaviors": [
+    "provider.merge-system-messages",
     "reasoning.strip-think-tags",
     "prompt.private-reasoning-guidance",
     "prompt.tool-cookbook-condensed",

package/data/chat-models/gl/glm-5.3-flash-320b-q2/manifest.json

@@ -28,8 +28,8 @@
     "name": "Z.ai",
     "url": "https://huggingface.co/zai-org/GLM-5.3-Flash"
   },
-  "version": "1.0.1",
-  "updatedAt": "2026-09-06T00:00:00Z",
+  "version": "1.0.2",
+  "updatedAt": "2026-10-17T00:00:00Z",
   "license": "MIT",
   "licenseClass": "open",
   "licenseShortName": "MIT",
@@ -130,8 +130,9 @@
       "sha256": "ae23e14c6979e889051b2e4a39351abcdafb161e18e606fae4d8c40095a4bf3a",
       "sizeBytes": 1127280960
     },
-    "residentBytes": 100865554432,
-    "ssdStreaming": false,
+    "residentBytes": 47244640256,
+    "cacheExpertsBytes": 34359738368,
+    "ssdStreaming": true,
     "maxLaunchCtx": 32768,
     "mtp": {
       "exactSampling": true

package/data/chat-models/gl/glm-5.3-flash-320b-q2/versions/1.0.2/manifest.json

@@ -0,0 +1,27 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.2",
+  "releasedAt": "2026-10-17T00:00:00Z",
+  "minGezelVersion": "1.26248",
+  "approxSizeBytes": 97633097344,
+  "ds4": {
+    "huggingfaceRepo": "antirez/glm-5.3-flash-gguf",
+    "revision": "b2fa29d7a6b410db11221c904973967b80b760f5",
+    "filename": "GLM-5.3-Flash-Q2.gguf",
+    "sha256": "e81fd6241c6e55a64e1e14e47a3eab61a173fa8d7e4b5c1d1848827119705b32",
+    "approxSizeBytes": 96505816384,
+    "quantization": "IQ2_XXS/Q2_K",
+    "visionEncoder": {
+      "filename": "GLM-5.3-Flash-Vision-Encoder.gguf",
+      "sha256": "ae23e14c6979e889051b2e4a39351abcdafb161e18e606fae4d8c40095a4bf3a",
+      "sizeBytes": 1127280960
+    },
+    "residentBytes": 47244640256,
+    "cacheExpertsBytes": 34359738368,
+    "ssdStreaming": true,
+    "maxLaunchCtx": 32768,
+    "mtp": {
+      "exactSampling": true
+    }
+  }
+}