MiniCPM-V-4_5

TWO-PART MODEL, and the small-but-capable positioning is a deliberate architecture choice, not just a size claim. (1) Vision: SigLIP, 27 layers, hidden 1152, 16 heads, patch 14, image_size 980 (note: 980, not 384 or 896 -- a large native input), batch_vision_input true. TILING + RESAMPLER: slice_mode true, slice_config.max_slice_nums 9, patch_size 14, and the decisive field query_num 64 -- a PERCEIVER-STYLE RESAMPLER that compresses every 448 slice down to 64 tokens (image_size 448 at top level, batch_3d_resampler true for video). VISION TOKENS: 64 per slice + 64 for the global view, so 128 tokens for a small image and up to 10 * 64 = 640 tokens at max_slice_nums 9. DOMINANCE: TEXT DOMINATES at default settings (128-640 vision tokens vs a text prompt), and vision stays bounded even at max tiling because the resampler caps per-slice cost. That bound is the whole design: MiniCPM pays for resolution inside the ViT (batch_vision_input) but caps what reaches the LLM. Cheapest vision-token path in this slice among the high-capability models. (2) LLM: dense decoder, 36 layers, hidden 4096, 32 heads / 8 KV heads, head_dim 128 (stated explicitly), intermediate 12288, vocab 151,748, max_position 40,960, rope_theta 1e6, use_sliding_window false. ARCHITECTURE = dense. No positional scaling applied (rope_scaling null). KV CACHE: 36 * 8 * 128 * 2 * 2 = 147,456 B/token (144 KiB). At 40,960 ctx that is 5.6 GiB per sequence at bf16 -- a far more honest serving figure than the 131k-262k contexts some siblings record, because 40,960 is what this model actually declares. params_b 8.696 = HF API safetensors total (8,695,895,280). 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]].