abstractcore 2.5.0__py3-none-any.whl → 2.5.3__py3-none-any.whl

abstractcore/assets/model_capabilities.json

@@ -13,6 +13,25 @@
       "aliases": [],
       "max_tokens": 128000
     },
+    "gpt-4-turbo": {
+      "max_output_tokens": 4096,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "max_tools": -1,
+      "vision_support": true,
+      "audio_support": false,
+      "image_resolutions": [
+        "variable"
+      ],
+      "notes": "GPT-4 Turbo with vision capabilities",
+      "source": "OpenAI official docs 2025",
+      "canonical_name": "gpt-4-turbo",
+      "aliases": [
+        "gpt-4-turbo-preview"
+      ],
+      "max_tokens": 128000
+    },
     "gpt-4-turbo-with-vision": {
       "max_output_tokens": 4096,
       "tool_support": "native",
@@ -45,6 +64,18 @@
       "image_resolutions": [
         "variable"
       ],
+      "image_tokenization_method": "tile_based",
+      "base_image_tokens": 85,
+      "tokens_per_tile": 170,
+      "tile_size": "512x512",
+      "max_image_dimension": 2048,
+      "short_side_resize_target": 768,
+      "detail_levels": [
+        "low",
+        "high",
+        "auto"
+      ],
+      "low_detail_tokens": 85,
       "notes": "Multimodal omni model, 2x faster, half price, 5x higher rate limits (updated Nov 2024)",
       "source": "OpenAI official docs 2025",
       "canonical_name": "gpt-4o",
@@ -154,6 +185,12 @@
       "image_resolutions": [
         "up to 1568x1568"
       ],
+      "image_tokenization_method": "pixel_area_based",
+      "token_formula": "(width * height) / 750",
+      "pixel_divisor": 750,
+      "max_image_dimension": 1568,
+      "token_cap": 1600,
+      "min_dimension_warning": 200,
       "audio_support": false,
       "notes": "disable_parallel_tool_use option available",
       "source": "Anthropic official docs",
@@ -316,7 +353,7 @@
     "llama-3.2-1b": {
       "max_output_tokens": 2048,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -329,7 +366,7 @@
     "llama-3.2-3b": {
       "max_output_tokens": 2048,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -342,7 +379,7 @@
     "llama-3.2-11b-vision": {
       "max_output_tokens": 2048,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": true,
       "image_resolutions": [
@@ -358,7 +395,7 @@
     "llama-3.3-70b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": true,
       "vision_support": false,
       "audio_support": false,
@@ -371,7 +408,7 @@
     "llama-3.1-8b": {
       "max_output_tokens": 8192,
       "tool_support": "native",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": true,
       "vision_support": false,
       "audio_support": false,
@@ -384,7 +421,7 @@
     "llama-3.1-70b": {
       "max_output_tokens": 8192,
       "tool_support": "native",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": true,
       "vision_support": false,
       "audio_support": false,
@@ -397,7 +434,7 @@
     "llama-3.1-405b": {
       "max_output_tokens": 8192,
       "tool_support": "native",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": true,
       "vision_support": false,
       "audio_support": false,
@@ -426,7 +463,7 @@
     "qwen2.5-0.5b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -439,7 +476,7 @@
     "qwen2.5-1.5b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -452,7 +489,7 @@
     "qwen2.5-3b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -465,7 +502,7 @@
     "qwen2.5-7b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -478,7 +515,7 @@
     "qwen2.5-14b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -491,7 +528,7 @@
     "qwen2.5-32b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -504,7 +541,7 @@
     "qwen2.5-72b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -517,7 +554,7 @@
     "qwen3-0.6b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -531,7 +568,7 @@
     "qwen3-1.7b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -545,7 +582,7 @@
     "qwen3-4b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -559,7 +596,7 @@
     "qwen3-32b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -573,7 +610,7 @@
     "qwen3-30b-a3b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -584,10 +621,26 @@
       "aliases": [],
       "max_tokens": 40960
     },
+    "qwen3-30b-a3b-2507": {
+      "max_output_tokens": 8192,
+      "tool_support": "prompted",
+      "structured_output": "native",
+      "parallel_tools": false,
+      "vision_support": false,
+      "audio_support": false,
+      "thinking_support": true,
+      "notes": "Qwen3-30B-A3B-Instruct-2507 with enhanced reasoning, coding, and mathematical skills. Supports up to 256K context, extendable to 1M tokens",
+      "source": "Alibaba Qwen3 2507 release",
+      "canonical_name": "qwen3-30b-a3b-2507",
+      "aliases": [
+        "qwen/qwen3-30b-a3b-2507"
+      ],
+      "max_tokens": 262144
+    },
     "qwen3-coder-30b": {
       "max_output_tokens": 8192,
       "tool_support": "native",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": true,
       "vision_support": false,
       "audio_support": false,
@@ -600,7 +653,7 @@
     "qwen2-vl": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": true,
       "image_resolutions": [
@@ -643,7 +696,7 @@
     "phi-3-mini": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -656,7 +709,7 @@
     "phi-3-small": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -668,7 +721,7 @@
     "phi-3-medium": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -680,7 +733,7 @@
     "phi-3.5-mini": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -692,7 +745,7 @@
     "phi-3.5-moe": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -705,7 +758,7 @@
     "phi-3-vision": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": true,
       "image_resolutions": [
@@ -720,7 +773,7 @@
     "phi-4": {
       "max_output_tokens": 16000,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -733,7 +786,7 @@
     "mistral-7b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -877,6 +930,31 @@
       "max_image_resolution": "1120x1120",
       "image_patch_size": 14,
       "max_image_tokens": 6400,
+      "image_tokenization_method": "resolution_tier_based",
+      "supported_resolutions": [
+        [
+          560,
+          560
+        ],
+        [
+          1120,
+          560
+        ],
+        [
+          560,
+          1120
+        ],
+        [
+          1120,
+          1120
+        ]
+      ],
+      "base_tokens_per_resolution": {
+        "560x560": 1600,
+        "1120x560": 3200,
+        "560x1120": 3200,
+        "1120x1120": 6400
+      },
       "notes": "Llama 3.2 Vision 11B model with multimodal capabilities for visual recognition and reasoning",
       "source": "Meta AI Llama 3.2 release",
       "canonical_name": "llama3.2-vision:11b",
@@ -941,7 +1019,7 @@
     "gemma-2b": {
       "max_output_tokens": 8192,
       "tool_support": "none",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -953,7 +1031,7 @@
     "gemma-7b": {
       "max_output_tokens": 8192,
       "tool_support": "none",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -1002,7 +1080,7 @@
     "codegemma": {
       "max_output_tokens": 8192,
       "tool_support": "none",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -1033,7 +1111,7 @@
     "glm-4": {
       "max_output_tokens": 4096,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -1085,7 +1163,7 @@
     "qwen3": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
-      "structured_output": "prompted",
+      "structured_output": "native",
       "parallel_tools": false,
       "vision_support": false,
       "audio_support": false,
@@ -1255,13 +1333,18 @@
       "video_support": true,
       "audio_support": false,
       "image_resolutions": [
-        "variable"
+        "64x64 to 4096x4096"
       ],
-      "max_image_resolution": "variable",
+      "max_image_resolution": "4096x4096",
       "image_patch_size": 16,
       "max_image_tokens": 24576,
       "pixel_grouping": "32x32",
-      "notes": "Qwen3-VL 4B dense model with 256K context, optimized for LMStudio",
+      "image_tokenization_method": "patch_based_adaptive",
+      "adaptive_resolution": true,
+      "min_resolution": 64,
+      "max_resolution": 4096,
+      "vision_encoder": "ViT-based",
+      "notes": "Qwen3-VL 4B dense model with 256K context, optimized for LMStudio. Parameters: 4.83B. FP8 checkpoints available.",
       "source": "Alibaba Qwen3-VL technical report 2025",
       "canonical_name": "qwen3-vl-4b",
       "aliases": [
@@ -1278,13 +1361,18 @@
       "video_support": true,
       "audio_support": false,
       "image_resolutions": [
-        "variable"
+        "64x64 to 4096x4096"
       ],
-      "max_image_resolution": "variable",
+      "max_image_resolution": "4096x4096",
       "image_patch_size": 16,
       "max_image_tokens": 24576,
       "pixel_grouping": "32x32",
-      "notes": "Qwen3-VL 8B dense model with 256K context, optimized for LMStudio",
+      "image_tokenization_method": "patch_based_adaptive",
+      "adaptive_resolution": true,
+      "min_resolution": 64,
+      "max_resolution": 4096,
+      "vision_encoder": "ViT-based",
+      "notes": "Qwen3-VL 8B dense model with 256K context, optimized for LMStudio. Parameters: 8.77B. FP8 checkpoints available.",
       "source": "Alibaba Qwen3-VL technical report 2025",
       "canonical_name": "qwen3-vl-8b",
       "aliases": [
@@ -1301,19 +1389,24 @@
       "video_support": true,
       "audio_support": false,
       "image_resolutions": [
-        "variable"
+        "64x64 to 4096x4096"
       ],
-      "max_image_resolution": "variable",
+      "max_image_resolution": "4096x4096",
       "image_patch_size": 16,
       "max_image_tokens": 24576,
       "pixel_grouping": "32x32",
-      "notes": "Qwen3-VL 30B MoE model (30.5B total/3.3B active), best performing vision model, 256K context",
+      "image_tokenization_method": "patch_based_adaptive",
+      "adaptive_resolution": true,
+      "min_resolution": 64,
+      "max_resolution": 4096,
+      "vision_encoder": "ViT-based",
+      "notes": "Qwen3-VL 30B MoE model (30.5B total/3.3B active), best performing vision model, 128K context",
       "source": "Alibaba Qwen3-VL technical report 2025",
       "canonical_name": "qwen3-vl-30b",
       "aliases": [
         "qwen/qwen3-vl-30b"
       ],
-      "max_tokens": 262144
+      "max_tokens": 131072
     },
     "qwen2.5-vl-7b": {
       "max_output_tokens": 8192,
@@ -1329,6 +1422,11 @@
       "image_patch_size": 14,
       "max_image_tokens": 16384,
       "pixel_grouping": "28x28",
+      "image_tokenization_method": "patch_based_adaptive",
+      "adaptive_resolution": true,
+      "min_resolution": 56,
+      "max_resolution": 3584,
+      "vision_encoder": "ViT-based",
       "notes": "Qwen2.5-VL 7B parameter vision model, 28x28 pixel patches, max 3584x3584 resolution",
       "source": "Alibaba official docs",
       "canonical_name": "qwen2.5-vl-7b",
@@ -1353,6 +1451,12 @@
       "vision_encoder": "SigLIP-400M",
       "image_tokens_per_image": 256,
       "adaptive_windowing": true,
+      "image_tokenization_method": "fixed_resolution",
+      "fixed_resolution": [
+        896,
+        896
+      ],
+      "preprocessing": "automatic_resize_and_crop",
       "notes": "Gemma3 4B parameter model with vision support, 896x896 fixed resolution with adaptive windowing",
       "source": "Google Gemma3 documentation 2025",
       "canonical_name": "gemma3-4b",
@@ -1547,6 +1651,7 @@
       "max_image_resolution": "768x768",
       "vision_encoder": "SigLIP2-so400m-patch14-384",
       "image_patch_size": 14,
+      "image_tokenization_method": "patch_based",
       "notes": "IBM Granite 3.2-Vision 2B model with SigLIP2 encoder, optimized for visual document understanding",
       "source": "IBM Granite 3.2 technical report arXiv:2502.09927",
       "canonical_name": "granite3.2-vision:2b",
@@ -1558,6 +1663,58 @@
       ],
       "max_tokens": 32768
     },
+    "gemini-2.5-flash": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "max_tools": -1,
+      "vision_support": true,
+      "audio_support": true,
+      "video_support": true,
+      "image_resolutions": [
+        "224x224",
+        "448x448",
+        "1024x1024"
+      ],
+      "max_image_resolution": "768x768",
+      "image_tokenization_method": "gemini_vision_encoder",
+      "thinking_support": true,
+      "thinking_budget": true,
+      "notes": "Optimized for speed and efficiency, suitable for high-volume, latency-sensitive tasks. Supports configurable thinking budgets",
+      "source": "Google AI official docs 2025",
+      "canonical_name": "gemini-2.5-flash",
+      "aliases": [
+        "gemini-2.5-flash-001"
+      ],
+      "max_tokens": 1000000
+    },
+    "gemini-2.5-pro": {
+      "max_output_tokens": 65536,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "max_tools": -1,
+      "vision_support": true,
+      "audio_support": true,
+      "video_support": true,
+      "image_resolutions": [
+        "224x224",
+        "448x448",
+        "1024x1024"
+      ],
+      "max_image_resolution":  "768x768",
+      "image_tokenization_method": "gemini_vision_encoder",
+      "thinking_support": true,
+      "thinking_budget": true,
+      "notes": "Most advanced Gemini model for complex reasoning, coding, and mathematical problem-solving. Features Deep Think mode for enhanced reasoning",
+      "source": "Google AI official docs 2025",
+      "canonical_name": "gemini-2.5-pro",
+      "aliases": [
+        "gemini-2.5-pro-001"
+      ],
+      "max_tokens": 1048576
+    },
     "granite3.3:2b": {
       "max_output_tokens": 8192,
       "tool_support": "prompted",
@@ -1587,6 +1744,321 @@
         "granite3.3-8b"
       ],
       "max_tokens": 32768
+    },
+    "embeddinggemma:300m": {
+      "max_output_tokens": 0,
+      "tool_support": "none",
+      "structured_output": "none",
+      "parallel_tools": false,
+      "vision_support": false,
+      "audio_support": false,
+      "notes": "Text embedding model, not for generation or vision",
+      "source": "Google Gemma documentation",
+      "canonical_name": "embeddinggemma:300m",
+      "aliases": [
+        "google/embeddinggemma-300m"
+      ],
+      "max_tokens": 0,
+      "model_type": "embedding"
+    },
+    "blip-image-captioning-base": {
+      "max_output_tokens": 512,
+      "tool_support": "none",
+      "structured_output": "none",
+      "parallel_tools": false,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "224x224",
+        "384x384"
+      ],
+      "max_image_resolution": "384x384",
+      "vision_encoder": "ViT-B/16",
+      "image_patch_size": 16,
+      "image_tokenization_method": "patch_based",
+      "base_image_tokens": 577,
+      "notes": "Salesforce BLIP image captioning model, primarily for image-to-text tasks",
+      "source": "Salesforce BLIP documentation",
+      "canonical_name": "blip-image-captioning-base",
+      "aliases": [
+        "Salesforce/blip-image-captioning-base"
+      ],
+      "max_tokens": 512
+    },
+    "glyph": {
+      "max_output_tokens": 8192,
+      "tool_support": "prompted",
+      "structured_output": "prompted",
+      "parallel_tools": false,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "variable"
+      ],
+      "max_image_resolution": "variable",
+      "base_model": "GLM-4.1V-9B-Base",
+      "total_parameters": "10B",
+      "tensor_type": "BF16",
+      "image_tokenization_method": "visual_text_compression",
+      "optimized_for_glyph": true,
+      "text_image_processing": true,
+      "architecture": "glm4v",
+      "requires_processor": true,
+      "message_format": "glm_special_tokens",
+      "conversation_template": {
+        "system_prefix": "<|system|>\n",
+        "system_suffix": "\n",
+        "user_prefix": "<|user|>\n", 
+        "user_suffix": "\n",
+        "assistant_prefix": "<|assistant|>\n",
+        "assistant_suffix": "\n"
+      },
+      "model_class": "AutoModelForImageTextToText",
+      "processor_class": "AutoProcessor",
+      "trust_remote_code": true,
+      "transformers_version_min": "4.57.1",
+      "notes": "Glyph framework for scaling context windows via visual-text compression. Built on GLM-4.1V-9B-Base. Renders long text into images for VLM processing. Requires AutoModelForImageTextToText and AutoProcessor with trust_remote_code=True.",
+      "source": "HuggingFace zai-org/Glyph model card",
+      "canonical_name": "glyph",
+      "aliases": [
+        "zai-org/Glyph"
+      ],
+      "max_tokens": 131072,
+      "license": "MIT",
+      "arxiv": "2510.17800",
+      "repository": "https://github.com/thu-coai/Glyph"
+    },
+    "glm-4.1v-9b-base": {
+      "max_output_tokens": 8192,
+      "tool_support": "prompted",
+      "structured_output": "prompted",
+      "parallel_tools": false,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "variable"
+      ],
+      "max_image_resolution": "variable",
+      "total_parameters": "9B",
+      "base_model": "GLM-4-9B-0414",
+      "image_tokenization_method": "glm_vision_encoder",
+      "architecture": "glm4v",
+      "requires_processor": true,
+      "message_format": "glm_special_tokens",
+      "model_class": "AutoModelForImageTextToText",
+      "processor_class": "AutoProcessor",
+      "trust_remote_code": true,
+      "transformers_version_min": "4.57.1",
+      "notes": "GLM-4.1V 9B base model, backbone for Glyph visual-text compression framework",
+      "source": "HuggingFace zai-org/GLM-4.1V-9B-Base",
+      "canonical_name": "glm-4.1v-9b-base",
+      "aliases": [
+        "zai-org/GLM-4.1V-9B-Base"
+      ],
+      "max_tokens": 131072
+    },
+    "glm-4.1v-9b-thinking": {
+      "max_output_tokens": 8192,
+      "tool_support": "prompted",
+      "structured_output": "prompted",
+      "parallel_tools": false,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 4096x4096"
+      ],
+      "max_image_resolution": "4096x4096",
+      "total_parameters": "10B",
+      "base_model": "GLM-4-9B-0414",
+      "image_tokenization_method": "glm_vision_encoder",
+      "thinking_support": true,
+      "reasoning_paradigm": "chain_of_thought",
+      "adaptive_resolution": true,
+      "aspect_ratio_support": "arbitrary",
+      "architecture": "glm4v",
+      "requires_processor": true,
+      "message_format": "glm_special_tokens",
+      "model_class": "AutoModelForImageTextToText",
+      "processor_class": "AutoProcessor",
+      "trust_remote_code": true,
+      "transformers_version_min": "4.57.1",
+      "notes": "GLM-4.1V-9B-Thinking with Chain-of-Thought reasoning, 64K context, arbitrary aspect ratios up to 4K resolution. First reasoning-focused VLM in the series, matches 72B models on 18 benchmark tasks.",
+      "source": "HuggingFace zai-org/GLM-4.1V-9B-Thinking and GitHub zai-org/GLM-V",
+      "canonical_name": "glm-4.1v-9b-thinking",
+      "aliases": [
+        "zai-org/GLM-4.1V-9B-Thinking",
+        "glm-4.1v-thinking",
+        "glm4.1v-9b-thinking"
+      ],
+      "max_tokens": 65536,
+      "arxiv": "2507.01006"
+    },
+    "mistral-small-3.1-24b-instruct": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 2048x2048"
+      ],
+      "max_image_resolution": "2048x2048",
+      "image_tokenization_method": "mistral_vision_encoder",
+      "notes": "Mistral Small 3.1 with 24B parameters, 128K context, multimodal understanding. Released March 2025.",
+      "source": "Mistral AI documentation and HuggingFace",
+      "canonical_name": "mistral-small-3.1-24b-instruct",
+      "aliases": [
+        "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
+      ],
+      "max_tokens": 131072,
+      "total_parameters": "24B",
+      "release_date": "2025-03-17"
+    },
+    "mistral-small-3.2-24b-instruct": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": true,
+      "audio_support": false,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 2048x2048"
+      ],
+      "max_image_resolution": "2048x2048",
+      "image_tokenization_method": "mistral_vision_encoder",
+      "tensor_type": "BF16",
+      "gpu_memory_required": "55GB",
+      "notes": "Mistral Small 3.2 with 24B parameters, 128K context. Improved instruction following, reduced repetition, enhanced function calling. Released June 2025.",
+      "source": "HuggingFace mistralai/Mistral-Small-3.2-24B-Instruct-2506",
+      "canonical_name": "mistral-small-3.2-24b-instruct",
+      "aliases": [
+        "mistralai/Mistral-Small-3.2-24B-Instruct-2506"
+      ],
+      "max_tokens": 131072,
+      "total_parameters": "24B",
+      "release_date": "2025-06-01"
+    },
+    "llama-4-scout": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": true,
+      "audio_support": true,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 1120x1120"
+      ],
+      "max_image_resolution": "1120x1120",
+      "architecture": "mixture_of_experts",
+      "active_parameters": "17B",
+      "total_parameters": "109B",
+      "experts": 16,
+      "image_tokenization_method": "resolution_tier_based",
+      "notes": "LLaMA 4 Scout with MoE architecture, 17B active/109B total parameters, 10M context window. Multimodal with early fusion. Released April 2025.",
+      "source": "Meta LLaMA 4 documentation and NVIDIA docs",
+      "canonical_name": "llama-4-scout",
+      "aliases": [
+        "llama4-17b-scout-16e-instruct",
+        "llama-4-17b-scout-16e-instruct"
+      ],
+      "max_tokens": 10000000,
+      "release_date": "2025-04-05",
+      "image_patch_size": 14,
+      "max_image_tokens": 6400
+    },
+    "llama-4-maverick": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": true,
+      "audio_support": true,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 1120x1120"
+      ],
+      "max_image_resolution": "1120x1120",
+      "architecture": "mixture_of_experts",
+      "active_parameters": "17B",
+      "total_parameters": "400B",
+      "experts": 128,
+      "image_tokenization_method": "resolution_tier_based",
+      "notes": "LLaMA 4 Maverick with MoE architecture, 17B active/400B total parameters, 1M context window. Optimized for coding and reasoning. Released April 2025.",
+      "source": "Meta LLaMA 4 documentation and Oracle docs",
+      "canonical_name": "llama-4-maverick",
+      "aliases": [
+        "llama4-17b-maverick-128e-instruct"
+      ],
+      "max_tokens": 1000000,
+      "release_date": "2025-04-05",
+      "image_patch_size": 14,
+      "max_image_tokens": 6400
+    },
+    "llama-4-behemoth": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": true,
+      "audio_support": true,
+      "video_support": false,
+      "image_resolutions": [
+        "up to 1120x1120"
+      ],
+      "max_image_resolution": "1120x1120",
+      "architecture": "mixture_of_experts",
+      "active_parameters": "288B",
+      "total_parameters": "2T",
+      "experts": 16,
+      "image_tokenization_method": "resolution_tier_based",
+      "notes": "LLaMA 4 Behemoth teacher model with 288B active/2T total parameters. Designed for distilling performance into smaller models. Announced April 2025.",
+      "source": "Meta LLaMA 4 announcement and PromptHub",
+      "canonical_name": "llama-4-behemoth",
+      "aliases": [
+        "llama4-288b-behemoth-16e"
+      ],
+      "max_tokens": 1000000,
+      "release_date": "2025-04-05",
+      "status": "announced",
+      "image_patch_size": 14,
+      "max_image_tokens": 6400
+    },
+    "minimax-m2": {
+      "max_output_tokens": 8192,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "vision_support": false,
+      "audio_support": false,
+      "video_support": false,
+      "thinking_support": true,
+      "architecture": "mixture_of_experts",
+      "active_parameters": "10B",
+      "total_parameters": "230B",
+      "thinking_paradigm": "interleaved_thinking",
+      "thinking_format": "<think>...</think>",
+      "notes": "MiniMax M2 MoE model optimized for coding and agentic workflows. Industry-leading 204K token context window. Uses interleaved thinking with <think> tags for reasoning. 10B active parameters from 230B total. Achieves strong performance on SWE-Bench and Terminal-Bench tasks. Supports complete tool calling for agent workflows.",
+      "source": "MiniMax official docs (minimax-m2.org, HuggingFace, GitHub)",
+      "canonical_name": "minimax-m2",
+      "aliases": [
+        "MiniMaxAI/MiniMax-M2",
+        "mlx-community/minimax-m2",
+        "mlx-community/MiniMax-M2",
+        "unsloth/MiniMax-M2-GGUF",
+        "minimax-m2-230b",
+        "minimax-m2-10b-active"
+      ],
+      "max_tokens": 208896,
+      "release_date": "2025-01",
+      "license": "Apache-2.0"
     }
   },
   "tool_support_levels": {
@@ -1605,6 +2077,73 @@
     "video_support": "Video processing capabilities",
     "fim_support": "Fill-in-the-middle code completion"
   },
+  "vlm_tokenization_research": {
+    "openai_gpt4v_formula": {
+      "step1": "Resize to fit 2048x2048 (preserve aspect ratio)",
+      "step2": "Resize shortest side to 768px",
+      "step3": "Calculate tiles: ceil(width/512) * ceil(height/512)",
+      "step4": "Total tokens = 85 + (tiles * 170)",
+      "low_detail": "Fixed 85 tokens regardless of size",
+      "research_source": "OpenAI official documentation + Image Tokenization research"
+    },
+    "anthropic_claude_formula": {
+      "formula": "min((width * height) / 750, 1600)",
+      "pixel_divisor": 750,
+      "token_cap": 1600,
+      "resize_trigger": "max(width, height) > 1568",
+      "warning_threshold": "min(width, height) < 200",
+      "research_source": "Anthropic Claude documentation + research analysis"
+    },
+    "google_gemini_formula": {
+      "small_image": "width <= 384 AND height <= 384 \u2192 258 tokens",
+      "large_image": "ceil(width/768) * ceil(height/768) * 258 tokens",
+      "small_threshold": 384,
+      "tile_size": 768,
+      "tokens_per_tile": 258,
+      "research_source": "Google Gemini documentation + research analysis"
+    },
+    "qwen_vl_adaptive_formula": {
+      "formula": "min(ceil(width/patch_size) * ceil(height/patch_size), max_tokens)",
+      "adaptive_resize": "Resize to fit within [min_res, max_res] range",
+      "patch_sizes": {
+        "qwen2.5": 14,
+        "qwen3": 16
+      },
+      "research_source": "Qwen-VL technical documentation + research"
+    },
+    "vision_transformer_baseline": {
+      "standard_patch_size": 16,
+      "formula": "tokens = (height * width) / (patch_size^2)",
+      "typical_range": [
+        196,
+        2048
+      ],
+      "research_source": "Vision Transformer foundational paper"
+    }
+  },
+  "generic_vision_model": {
+    "max_output_tokens": 4096,
+    "tool_support": "prompted",
+    "structured_output": "prompted",
+    "parallel_tools": false,
+    "vision_support": true,
+    "audio_support": false,
+    "video_support": false,
+    "image_resolutions": [
+      "up to 1024x1024"
+    ],
+    "max_image_resolution": "1024x1024",
+    "image_patch_size": 16,
+    "max_image_tokens": 2048,
+    "image_tokenization_method": "patch_based",
+    "adaptive_resolution": false,
+    "vision_encoder": "generic_vit",
+    "notes": "Generic vision model fallback with conservative parameters that should work with most VLMs",
+    "source": "AbstractCore generic fallback",
+    "canonical_name": "generic_vision_model",
+    "aliases": [],
+    "max_tokens": 32768
+  },
   "default_capabilities": {
     "max_output_tokens": 4096,
     "tool_support": "none",