On August 31, DeepSeek put the weights of DeepSeek-V4-Flash-Vision-Exp — the first experimental multimodal model in the V4 family — on Hugging Face, just 10 days after its API launch on August 21. 305B parameters, MIT license, and a minimal PyTorch inference reference implementation. Now that the weights are open, the most useful thing to do is reread the official 11-benchmark table item by item: which wins are real, and which are not.

Three wins over Opus — but not Chartography

The model card's three-way comparison (Vision-Exp vs. the previous 0731 text-only release vs. Anthropic Opus-4.8) covers 7 text-agent benchmarks and 4 multimodal ones. Checked line by line, Vision-Exp actually beats Opus-4.8 on exactly three: DeepSWE 59.3 vs 58.0 on the text side; ZeroBench (Pass@5) 35.0 vs 34.0 and Agents' Last Exam 27.3 vs 25.7 on the multimodal side.

The often-cited Chartography, however, is 64.3 vs 65.0 — Opus ahead by 0.7. ApexBench (Pass@1) is 36.5 vs 39.4, also Opus. On the text side, Terminal Bench 2.1 (83.9 vs 85.0), NL2Repo (57.7 vs 69.7), Cybergym (75.3 vs 78.3), DSBench-Hard (63.6 vs 71.7) and AutomationBench (25.7 vs 27.2) all favor Opus too. The pattern is structural: Opus still leads on heavy text tasks, while Vision-Exp closes in — and occasionally overtakes — on vision-related and SWE ground.

Against the 0731 predecessor, six of seven text benchmarks improve: Terminal Bench 2.1 from 82.7 to 83.9, DeepSWE from 54.4 to 59.3, Toolathlon-Verified from 70.3 to 75.9, DSBench-Hard from 59.6 to 63.6, NL2Repo from 54.2 to 57.7, AutomationBench from 25.1 to 25.7; the single dip is Cybergym (76.7 to 75.3). The multimodal jump is larger: ApexBench from 26.2 to 36.5 (+10.3) — note the table footnote: 0731 ignored the multimodal elements in its input for this benchmark, so those 10 points are the direct payoff of adding eyes. Agents' Last Exam goes from 25.2 to 27.3.

What else is in the repository

The reference implementation covers the vision encoder, aligner, DFlash attention, MoE, Hyper-Connections, and the DSpark forward path — component names that map to DeepSeek's own stream of attention work this year, not a bolt-on of off-the-shelf modules. Evaluation configuration is explicit: text-agent benchmarks use the minimal mode of DeepSeek Harness with the max reasoning effort level, temperature = 1.0, top_p = 0.95.

Deployment recipes go down to the command line: a single vLLM docker command serves the model on one 4×GB300 node with fp8 KV cache, block size 256, tensor parallel 4, and DSpark speculative decoding on by default (num_speculative_tokens = 3). On SGLang, --speculative-algorithm DSPARK draws draft and target weights from the same checkpoint, so no separate draft model is needed. The ecosystem moved fast too: about two days in, Hugging Face shows 17,893 monthly downloads, 2 finetunes and 8 quantized variants, with llama.cpp, LM Studio and Ollama adapter entries already in place.

So what for developers

While the model is API-only, you have to take the vendor's scores on faith; with weights on disk, anyone can re-run the table under the same configuration — including verifying which benchmarks were actually won. The MIT license carries no commercial strings. The next signal worth watching: when the Exp suffix drops, that is most likely the ship date of the official V4 multimodal release.

Reference: huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp (model card and benchmark table); theopenweights.com/news/deepseek-v4-flash-vision-exp-2l4q