Pixtral-12B-2409

TWO-PART MODEL, and the cost profile here is the outlier in this slice. (1) Vision: ViT, 24 layers, hidden 1024, 16 heads, patch_size 16, image_size 1024, intermediate 4096, rope_theta 1e4. There is NO spatial merge: every 16x16 patch becomes one LLM token. (2) LLM: Mistral dense decoder, dim 5120, n_layers 40, n_heads 32 / n_kv_heads 8, head_dim 128, hidden_dim 14336, vocab 131,072, rope_theta 1e9, max_position 131,072. ARCHITECTURE = dense. ACTIVATION is null: params.json does not state it (not published in any machine-readable form). VISION TOKENS -- THE HEADLINE NUMBER: a 1024x1024 tile is (1024/16)^2 = 4,096 ViT patches = 4,096 LLM tokens, one token per patch with no pooling whatsoever. Pixtral splits large images into 1024x1024 tiles; a 4096x4096 image is a 4x4 grid = 16 tiles = 65,536 VISION TOKENS. (Mistral's processor caps the tile count via max_num_tiles, so a serving config will lower this -- but the per-tile cost of 4,096 tokens is fixed by the architecture.) DOMINANCE: VISION DOMINATES CATEGORICALLY. Even one 1024x1024 tile (4,096 tokens) is 4x a typical 1,000-token text prompt, and multi-tile inputs reach tens of thousands of tokens. Compare: the same image at 1024x1024 through InternVL3 (448 tiles + 2x2 pixel shuffle) costs ~2,048 tokens, so Pixtral pays ~2x the prefill of InternVL3 and ~16x Qwen2.5-VL (256 tokens at 448x448) for the same pixels. Pixtral is the worst vision-token-per-pixel model in this slice. KV CACHE: 40 * 8 * 128 * 2 * 2 = 163,840 B/token (160 KiB) -- the largest per-token KV of any VLM recorded here. At 131,072 ctx that is 20.0 GiB per sequence at bf16. params_b 12.0 is MISTRAL'S OWN FIGURE, NOT verified from tensors: the repo ships consolidated.safetensors with no HF config.json (404) and the API reports no safetensors parameter metadata. Flagged as a vendor claim; treat as +/-. params.json is Mistral's reference config and is the source for every dimension field. SERVICEABILITY JOIN (added 2026-10-03; the architecture x engine matrix this record was missing, with a citation per cell in docs/13-model-compatibility.md): FAMILY=multimodal. REQUIRED KERNEL CLASS: an encoder tower (ViT / SigLIP / ConvNet / audio encoder) running on a SEPARATE schedule from the decoder, plus a projector into the decoder's embedding space, plus - for the mrope and DeepStack families - a multi-axis rotary scheme in the decoder rather than plain 1-D RoPE. Three engine-side consequences, all cited. (1) The encoder tower is a separate compute path, so its throughput is a different problem from decode throughput and engines that disclaim encoder throughput are disclaiming this. vLLM says so in its own words: 'We currently support pooling models primarily for convenience. This is not guaranteed to provide any performance improvements over using Hugging Face Transformers or Sentence Transformers directly' [[vllm-pooling-models]] - that is the general statement this class inherits. TRT-LLM's answer is a documented mechanism: a Multimodal Encoder Side Stream that prefetches encoder work on a separate CUDA stream plus a per-model Multimodal Embeddings Cache, currently available only for seven architectures (Gemma4, Gemma4Unified, Mistral3, Qwen3VL, Qwen3VLMoe, Qwen3_5, Qwen3_5Moe) [[mdl2-trtllm-supported-models]]. (2) Several towers carry hard external dependencies: Gemma3nForConditionalGeneration 'depends on timm>=1.0.17 to make use of its MobileNet-v5 vision backbone' and 'is only supported on V1 due to shared KV caching' [[mdl2-vllm-supported-models-2026-10]]; MiniCPM-V 4.6 requires transformers>=5.7.0 because it was upstreamed as a native model type [[mdl2-trtllm-supported-models]]. (3) Hybrid-only multimodal models expose a memory knob unique to this class: vLLM's --language-model-only flag sets all supported multimodal modalities to 0 so the multimodal modules are not loaded, 'to free up more GPU memory for KV cache' for Llama-4, Step3, Mistral-3 and Qwen-3.5 [[mdl2-vllm-supported-models-2026-10]].SUPPORTED BY (cited): vLLM native rows include InternVLChatModel and InternVLForConditionalGeneration, LlavaOnevisionForConditionalGeneration, LlavaNextForConditionalGeneration, MolmoForCausalLM and Molmo2ForConditionalGeneration, PaliGemmaForConditionalGeneration, SmolVLMForConditionalGeneration, Glm4vForConditionalGeneration, Qwen2VL/Qwen3VL, MiniCPMV, Mllama, Phi4MM, Gemma4, NemotronH_Nano_VL_V2, VoxtralForConditionalGeneration (which requires mistral-common[audio]) and Qwen3OmniMoeThinkerForConditionalGeneration [[mdl2-vllm-supported-models-2026-10]]. SGLang's multimodal table covers InternLM2-VL, LLaVA-OneVision, LLaVA-NeXT(-Video), Ling-3.0-flash-VL, LFM2.5-VL, MiniCPM-V-2, GLM-4.5V, GLM-OCR, Qwen3-VL 235B-A22B and 30B-A3B, Qwen3-Omni-30B-A3B, Kimi-VL-A3B, MiMo-VL-7B-RL, Step3-VL-10B, dots.ocr, PaddleOCR-VL, Cosmos3, deepseek-vl2, Janus-Pro-7B, NVILA-8B and gemma-3-4b-it, plus openai/whisper-large-v3 and Qwen3-ASR-1.7B on the audio route [[mdl2-sglang-multimodal-language-models]]. TRT-LLM's multimodal matrix covers a narrower set with an explicit per-row modality string and per-row Untested cells [[mdl2-trtllm-supported-models]]. Older GGUF-side engines reach the class through llama.cpp's projector graph, which llama.cpp-embedding-server states it cannot serve on the embedding route [[llama-cpp-embedding-server]]. PER-MODALITY CAVEATS CITED: TRT-LLM marks Gemma4Unified and MiniCPM-V 4.6 Untested for CUDA Graph / chunked prefill, MiniMax-M3 Untested for attention DP and KV cache reuse, Qwen4Exp Untested across the board, and notes audio requires a checkpoint with a sound_config and is supported only on the full non-disaggregated path [[mdl2-trtllm-supported-models]]. Step-3.7's vision tower runs in BF16 even when the text decoder is quantized [[mdl2-trtllm-supported-models]] - a precision asymmetry that silently costs memory. vLLM's fallback path accepts image inputs only, with video 'to be added in a future release' [[mdl2-vllm-supported-models-2026-10]]. CONTEXT_SCALING_METHOD (native): rope_theta 1e9 with rope_scaling null - same trap as Magistral: 1e9 is the training base.