開放模型與多模態推論
DeepSeek-V4-Flash-Vision-Exp Releases 305B Weights, but vLLM Still Cannot Load Vision Tensors
Ten days after launching the API, DeepSeek released the weights, prompt encoder, and minimal PyTorch inference implementation for its first V4 multimodal model under the MIT License. Although the model card shows generic deployment instructions, vLLM community testing found that the model still cannot be loaded directly because 316 vision tensors have no corresponding modules.
DeepSeek has uploaded the 305B-parameter weights for **DeepSeek-V4-Flash-Vision-Exp** to Hugging Face, filling the gap left by the hosted-only API released on August 21. It is the first official multimodal model in the V4 series: DeepSeek added a vision encoder and alignment module to V4-Flash, then continued training it to process mixed image, text, and tool inputs. The weights and reference code are licensed under the MIT License.
The release includes more than a checkpoint. The repository contains an encoder that converts OpenAI-style messages into model prompts, along with a minimal PyTorch implementation covering the vision encoder, aligner, DFlash attention, MoE, Hyper-Connections, and DSpark forward path. It also ensures that JSON image content blocks and `<image>path</image>` notation produce the same token sequence, making it easier for inference frameworks to verify preprocessing results. The official API accepts base64-encoded images, external URLs, or the Files API, with each image billed at no more than 384 tokens.
DeepSeek’s internal evaluation shows that the model scored 36.5 Pass@1 on ApexBench, up from 26.2 for V4-Flash-0731 when image inputs were ignored; it achieved 35.0 Pass@5 on ZeroBench. However, text-agent evaluations did not improve across the board. For example, the Cybergym score fell from 76.7 to 75.3. Moreover, all results were obtained using DeepSeek Harness, the highest reasoning effort, and specified sampling parameters, so they cannot be directly generalized to other agent stacks.
Deployers should be especially careful not to treat the autogenerated `vllm serve` example on the Hugging Face page as proof of compatibility. A newly opened vLLM support issue reports that the current routing logic passes the checkpoint to a text-only class, leaving 316 vision tensors without matching destination modules. The next developments to watch are official vLLM and SGLang model integrations, VRAM and multi-GPU topology requirements, and whether third parties can reproduce the multimodal agent results.