huggingface.co web signal

DeepSeek releases 305B V4-Flash-Vision-Exp under MIT license

By Published Updated
4 sources tracking this story
DeepSeek Open Source Multimodal China AI ai-business

TL;DR

  • DeepSeek priced V4-Flash-Vision-Exp at the same rates as text-only V4-Flash, forcing a direct cost comparison against Western multimodal API pricing.
  • The 384-token-per-image billing cap is an architectural constraint of the vision encoder itself, not a commercial policy that DeepSeek can easily relax.
  • Open weights arrived ten days after the API launch, compressing DeepSeek's closed-to-open sequencing to a window short enough that developers treating it as an open model from day one face minimal lock-in risk.

DeepSeek published a 305B-parameter experimental multimodal model, DeepSeek-V4-Flash-Vision-Exp, to Hugging Face under an MIT license. It is the first vision model in the V4-Flash family, adding a vision encoder and aligner on top of the text model's DFlash attention, MoE, Hyper-Connections and DSpark forward path.

The model card is spare. Its lone evaluative sentence reads: "Compared to DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash-Vision-Exp achieves substantial improvements on its multimodal agent capabilities, while maintaining comparable performance on text-only agent tasks."

The benchmark table puts Vision-Exp at 36.5 on ApexBench Pass@1, up from 26.2 for the text-only V4-Flash-0731 and shy of Opus-4.8's 39.4. Agents' Last Exam scores 27.3, ahead of Opus's 25.7. ZeroBench Pass@5 lands at 35.0 against Opus's 34.0, and Chartography at 64.3 against 65.0. On text-only agent tasks the gap is narrower: Terminal Bench 2.1 at 83.9 versus Opus's 85.0, DeepSWE at 59.3 versus Opus's 58.0, and NL2Repo at 57.7 against Opus's 69.7.

Weights ship with `vllm serve` and SGLang launch commands baked into the README, a minimal PyTorch inference script, and quantizations flagged for llama.cpp, Ollama and LM Studio. No paper or blog post accompanied the drop.

It arrives two days after Tencent open-sourced Hy4-preview, continuing the pace on the China AI beat we track.

What others are reporting

Coverage cluster as of 24h after publish

  1. DeepSeek API Docs Read →

    First-party announcement confirming August 21 API launch, image tokenization at up to 384 tokens billed at V4-Flash rates, and simultaneous Harness 0.1.1 tooling release.

    V4-Flash-Vision-Exp matches DeepSeek-V4-Flash on text capabilities while making a major leap on multimodal agent benchmarks.
  2. NYU Shanghai RITS Read →

    Technical analysis arguing the 384-token image cap is an architectural limit, not policy; critiques benchmark methodology for inflating multimodal gains vs text-only baselines.

    The ten-day gap is now short enough that the distinction is about sequencing rather than commitment.
  3. Open Source For You Read →

    Developer-community framing: MIT license on Hugging Face positions this as DeepSeek's open-source commitment rather than a one-time release, inviting downstream fine-tuning.

    DeepSeek has made V4-Flash-Vision-Exp available through Hugging Face, giving developers and researchers a platform to access the model.

Shared on Bluesky by 1 AI expert

Originally reported by huggingface.co

Read the original article →

Original headline: DeepSeek Drops V4-Flash-Vision-Exp, a 305B MIT-Licensed Multimodal Agent Model