DeepSeek publishes 168GB vision model weights under an MIT license

DeepSeek released the 168GB weights for its 305B-parameter DeepSeek-V4-Flash-Vision-Exp model under an MIT license on August 31, 10 days after the API debut, enabling local deployment and modification of the first vision model in the V4 family. The model card reports a 27.3 score on Agents' Last Exam, ahead of Anthropic's Claude Opus 4.8 at 25.7, though the comparison uses a superseded version of Opus.

DeepSeek publishes 168GB vision model weights under an MIT license The 305B-parameter checkpoint lands 10 days after its API debut, while DeepSeek's headline benchmark uses Anthropic's superseded Opus 4.8. By Ryan Merket /author/ryan-merket · Published Primary source: ModelScope on X https://x.com/ModelScope2022/status/2094421481983734088 Why it matters DeepSeek has moved its first V4 vision model from an API product to a downloadable checkpoint, widening access to multimodal agent development while keeping the hardware bar high. DeepSeek founder Liang Wenfeng https://apnews.com/article/0673d5c39d90108189cc31b88d85b9f8 published open weights for DeepSeek-V4-Flash /models/azure/deepseek-v4-flash -Vision-Exp on August 31st, giving developers a downloadable version of the lab's first experimental multimodal model in the V4 family. The release is new, though the model is not. DeepSeek opened API access https://api-docs.deepseek.com/news/news260821/ on August 21st, 10 days before ModelScope disclosed the weights release in a post on X https://x.com/ModelScope2022/status/2094421481983734088 . The API launch added image inputs to a model family that had previously focused on text, reasoning and agent workflows. Monday's release puts the underlying checkpoint and reference inference code into developers' hands under the permissive MIT license. Liang founded DeepSeek in 2023 after building High-Flyer, the quantitative hedge fund that provided the financial and computing base for his AI work. DeepSeek has since used open weights as a distribution strategy, allowing outside developers to run and modify its models without sending every request through DeepSeek's servers. The vision release extends that approach to image-based agent tasks. A large model, now available for local deployment The official Hugging Face repository https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp lists DeepSeek-V4-Flash-Vision-Exp at 305 billion parameters. Its 168GB repository divides the weights across 48 Safetensors files and includes a tokenizer, prompt-encoding reference and a minimal PyTorch inference implementation. DeepSeek says the model builds on the V4 Flash architecture with added visual modules and continued training for image understanding. The reference implementation covers the vision encoder and aligner alongside DeepSeek's mixture-of-experts architecture, DFlash attention, Hyper-Connections and DSpark forward path. DeepSeek also provides instructions for serving the checkpoint with vLLM and SGLang. The 168GB download narrows the practical audience. Research groups, inference providers and developers with multi-GPU systems can run or adapt the weights independently. Most individual developers will still find DeepSeek's hosted API easier to use. Hugging Face listed no hosted inference provider for the checkpoint at publication. DeepSeek's vision API documentation https://api-docs.deepseek.com/guides/vision/ supports mixed image and text prompts through OpenAI-compatible Chat Completions and Responses interfaces, plus an Anthropic-compatible Messages endpoint. Developers can submit JPEG, PNG, GIF and WebP files as base64 data, public URLs or references to DeepSeek's Files API. DeepSeek caps image processing at 384 tokens per image after resizing. A request can contain as many as 600 images, subject to file-size and request limits. DeepSeek said at the API launch that those visual tokens would be charged at V4 Flash rates. The benchmark claim needs a date stamp DeepSeek's model card reports that V4 Flash Vision scored 27.3 on Agents' Last Exam, ahead of Claude Opus 4.8 /models/anthropic/claude-opus-4.8-fast at 25.7. It also reported a 35.0 pass-at-five score on ZeroBench, compared with 34.0 for Opus 4.8. On Chartography, DeepSeek reported 64.3 against Opus 4.8's 65.0. The complete table is less tidy than ModelScope's post. V4 Flash Vision trailed Opus 4.8 on ApexBench, 36.5 to 39.4, leaving DeepSeek ahead on two of the four multimodal agent evaluations and behind on two. Across seven text-agent tests, DeepSeek reported one win over Opus 4.8, on DeepSWE, and six losses. Those results are supplied by DeepSeek rather than an independent evaluator. DeepSeek ran its models through the minimal configuration of DeepSeek Harness with maximum reasoning effort, a temperature of 1.0 and top-p of 0.95. Its comparison with the text-only V4 Flash baseline also requires care: that baseline ignored visual elements on two multimodal tests. Opus 4.8 is no longer Anthropic's current Opus model. Anthropic released Claude Opus 5 https://www.anthropic.com/news/claude-opus-5 on July 24th and described it as an improvement over Opus 4.8. DeepSeek's table therefore measures its new open-weight vision model against a closed model that Anthropic replaced more than a month earlier. The weights release remains the consequential part of Monday's news. Developers can now inspect and deploy DeepSeek's multimodal V4 checkpoint on their own infrastructure, rather than treating its visual capabilities as an API-only service. The MIT license also gives model builders broad room to modify, fine-tune and incorporate the checkpoint into their own products, provided they can supply the hardware.