Loading page…
The release gives developers the model files and serving references needed to run or adapt DeepSeek-V4-Flash-Vision-Exp independently. Its 168GB size means that freedom is aimed more at infrastructure-equipped teams than typical individual users.
Listen to this story
DeepSeek has moved V4 Flash Vision Exp from hosted API toward self-hosting by releasing its 305B-parameter checkpoint, code, and serving guidance under MIT. The 168GB package broadens inspection and deployment options, but effectively targets multi-GPU operators rather than typical developers; no hosted Hugging Face inference provider was listed at publication. API access remains easier, with OpenAI- and Anthropic-compatible endpoints and up to 600 images per request.
API access began August 21; open weights and reference code followed 10 days later.
The repository contains 48 Safetensors files, a tokenizer, prompt-encoding reference, and PyTorch inference code.
DeepSeek caps each image at 384 processing tokens, despite allowing up to 600 images per request.
DeepSeek has published downloadable weights for its experimental V4 Flash vision model under the MIT license, moving the 305-billion-parameter system from API-only access toward self-hosted deployment. The shift gives developers more control over a model that accepts images, but the 168GB checkpoint sets a high infrastructure threshold for anyone seeking to run it locally.
DeepSeek-V4-Flash-Vision-Exp is based on the V4 Flash architecture, with visual modules added and further training for image understanding. The repository packages its weights in 48 Safetensors files, alongside a tokenizer, prompt-encoding reference and a minimal PyTorch inference implementation.
That combination is more than a file drop. The release also includes instructions for serving the checkpoint with vLLM and SGLang, two systems used to operate models as services. Teams can therefore inspect, deploy and adapt the supplied checkpoint rather than send every image-and-text request through DeepSeek’s hosted product.
A permissive license permits broad use, modification and incorporation of the checkpoint into other products. Yet a 168GB download is not a consumer-device deployment: the model is positioned for research groups, inference providers and developers with multi-GPU systems, while the hosted API will remain the simpler path for most individual developers.
The API accepts mixed image-and-text prompts through OpenAI-compatible Chat Completions and Responses interfaces, as well as an Anthropic-compatible Messages endpoint. Developers can provide JPEG, PNG, GIF or WebP images as base64 data, public URLs or references through DeepSeek’s Files API.
DeepSeek caps processing at 384 tokens per image after resizing, while a request can contain up to 600 images subject to file-size and request limits. Those numbers describe the API’s input boundaries, not a demonstrated measure of visual accuracy or reliability on a particular task.
DeepSeek’s model card reports scores of 27.3 on Agents’ Last Exam and 35.0 pass-at-five on ZeroBench, versus 25.7 and 34.0 for Claude Opus 4.8. The same table places V4 Flash Vision behind Opus 4.8 on ApexBench and Chartography, and across seven text-agent tests DeepSeek reported one win and six losses.
Those are company-reported results, run with DeepSeek Harness’s minimal configuration, maximum reasoning effort, temperature 1.0 and top-p 0.95. They also compare the new release with Claude Opus 4.8, not Anthropic’s newer Opus 5, which Anthropic released on July 24 as an improvement over Opus 4.8.
Editorial analysis
This is a distribution change more than a new vision launch: DeepSeek’s image-capable endpoint had already been available for 10 days, while the checkpoint now lets organizations move inference and modification onto their own infrastructure. The important test is whether serving providers add hosted access or whether teams with multi-GPU capacity begin adapting the model themselves. That divide will determine whether the MIT license creates broad practical access or chiefly more optionality for research groups and inference operators. DeepSeek’s earlier 384-token image cap also remains a useful constraint to watch in real multimodal workloads.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.