DeepSeek has published DeepSeek-V4-Flash-Vision-Exp on Hugging Face, the first multimodal variant in the V4 family, released under the MIT license. It adds a vision tower to the DeepSeek-V4-Flash architecture, with continued training for agent tasks involving images, charts and screenshots. On two multimodal benchmarks it scores above closed Opus-4.8.
Key takeaways
- 304.6B parameters, MoE with 256 routed experts, six active per token
- Context up to 1,048,576 tokens, extended from 65,536 via YaRN
- ApexBench Pass@1: 36.5 against 26.2 for the earlier 0731 release
- Agents' Last Exam: 27.3 points, above Opus-4.8 at 25.7
- MIT license, checkpoint roughly 168 GB across 48 shards
A vision tower built on Flash
The language core is an MoE of 256 routed experts?MoE routing: A router in the MoE layer picks a handful of experts from the full pool for every token, leaving the rest of the network idle. That is how a 304.6B-parameter model computes only a fraction of itself per token. plus one shared, six firing per token. The vision encoder is a separate tower bolted onto that core. From base Flash the model inherits DFlash attention, Hyper-Connections and DSpark, the built-in speculative decoding path.
Architecture at a glance:
Gains where the model has to look
The largest jump lands on ApexBench, but the stronger signal comes from Agents' Last Exam and ZeroBench: on both, the open model leads Opus-4.8.
| Model | ApexBench | ALE | ZeroBench | Terminal 2.1 | DeepSWE | Toolathlon |
|---|---|---|---|---|---|---|
| V4-Flash-Vision-Exp | 36.5 | 27.3 | 35.0 | 83.9 | 59.3 | 75.9 |
| V4-Flash-0731 | 26.2 | — | — | 82.7 | 54.4 | 70.3 |
| Opus-4.8 | — | 25.7 | 34.0 | — | — | — |
The vision training did not cost text performance — Terminal Bench 2.1, DeepSWE and Toolathlon-Verified all move up. The gap to Opus-4.8 stays wide on repository generation (57.7 against 69.7) and harder data analysis (63.6 against 71.7).
Hardware floor
Running it takes data-center hardware. The vLLM recipe lists one GB200 NVL4 tray as reference and roughly 202 GB of VRAM before KV cache allocation. Experts sit in FP4, remaining weights in FP8, the vision tower in BF16.
Why it matters
An open-weights model under the MIT license that matches the closed frontier on part of the multimodal workload shifts the reference point for teams building agents on their own infrastructure. What counts is not a single score but the combination: a permissive license, predictable inference cost and no vendor API dependency. The Exp suffix signals that DeepSeek treats this as a testbed, not a production-ready product.
What next?
- vLLM names NVIDIA as the only confirmed platform — ROCm and XPU builds lack vision support
- DeepSeek has published no technical report and no date for a stable multimodal release
Sources
- Hugging Face — DeepSeek-V4-Flash-Vision-Exp
- vLLM Recipes — DeepSeek-V4-Flash-Vision-Exp recipe
