What is DeepSeek V4-Flash-Vision-Exp 2026: open weights and evals
DeepSeek open-sourced its first V4 vision checkpoint. Self-reported scores split with Opus-4.8; no independent lab has reproduced the harness.
V4 could write code from the spring release onward. It could not look at a picture. On 31 August DeepSeek put the vision checkpoint’s weights on Hugging Face under MIT.
The query is DeepSeek V4-Flash-Vision-Exp: a ~305B-parameter sparse MoE whose text backbone is still V4-Flash-0731. What follows matches TechTimes (1 Sep 2026) against the model card and the lab’s table. Vendor numbers are not treated as independent replications.
What the weights add
The backbone remains a ~284B sparse MoE, ~13B active per step, 43 layers, hidden size 4,096, one-million-token context; routing picks 6 of 256 experts plus a shared expert. Vision is a 32-layer, 1,024-d, 16-head encoder on 14×14 patches, aligned into the 4,096-d space, capped at 384 tokens per image — the same cap as billing. The two modules add roughly 21B parameters.
How to read the self-report
- 1
The harness is closed
Comparisons to V4-Flash-0731 and Opus-4.8 used Harness Minimal Mode at max reasoning effort, temperature 1.0, top_p 0.95. The code is unpublished, so an outside lab cannot rerun the same config yet.
- 2
A split versus Opus-4.8
Three of eleven wins: DeepSWE 59.3 vs 58.0, Agents’ Last Exam 27.3 vs 25.7, ZeroBench Pass@5 35.0 vs 34.0. Terminal Bench 2.1, Toolathlon-Verified and Chartography sit within about a point. NL2Repo 57.7 vs 69.7 and DSBench-Hard 63.6 vs 71.7. A footnote: the text baseline ignores multimodal input on ApexBench and Agents’ Last Exam, so part of those leaps is structural.
- 3
Text did not collapse
DeepSeek says the vision variant “maintains comparable performance” on text-only agent tasks. The table has it ahead of V4-Flash-0731 on six of seven text benches (Toolathlon-Verified +5.6, DeepSWE +4.9). Cybergym is the sole regression versus the text baseline (75.3 vs 76.7). Precedent: yage.ai saw V4-Pro’s vendor SWE-bench Verified 80.6% fall to about 8% pass@1 on a stricter DeepSWE. The 59.3 DeepSWE figure uses the tighter suite — and is still self-reported.
A narrow eval surface
The lab measured agentic and chart-reading work, not broad VQA, OCR or science-vision suites. Useful for software pipelines; thin as a general “can it see” claim. Related open-weight notes: Qwen weights and Ornith.
Exp is not production-stable
DeepSeek uses Exp for checkpoints still under evaluation. The production text Flash is V4-Flash-0731. No announced date for a stable vision build, or for FP4 experts on that build.
Hosted and self-hosted are different risks
Prompts and images to api.deepseek.com transit servers in China. MIT weights on your own machines cut that outflow. No named independent security audit of these weights is cited as of publication. Same rule as our Guidelight note: quote a number only when a source printed it.
| Claim | Checkable source | How to read it |
|---|---|---|
| MIT ~305B vision MoE | Hugging Face card / TechTimes 2026-09-01 | Self-hostable; hardware is the gate |
| 3 of 11 leads vs Opus-4.8 | DeepSeek’s own harness table | Vendor-reported, not independently reproduced |
| API 21 Aug, weights 31 Aug | OpenRouter / card timeline | Ten-day telemetry window, unlike same-day 0731 |
TechTimes: whether the weights are worth the multi-GPU bill depends on the lab’s own harness — and no independent lab has reproduced it yet.
# DeepSeek-V4-Flash-Vision-Exp
# weights: 2026-08-31 API: 2026-08-21
id: deepseek-v4-flash-vision-exp
license: MIT
params: ~305B sparse MoE (vision tower + 284B text)
harness: unpublished Harness Minimal Mode
independent_repro: none as of 2026-09-02Boundaries
- MIT ≠ a single-GPU toy
- The license allows commercial use and fine-tunes. VRAM and multi-GPU setups are the real gate. The card’s example includes vLLM on 4×GB300.
- A vendor lead ≠ a general vision crown
- The suite leans software agents and charts. Long-horizon synthesis still trails Opus-4.8.
- This page is not a meeting-product brief
- It only covers the release and the eval. For a short shared board, open wbmeet and create or join a room.
Questions worth checking
How does it differ from text V4-Flash-0731?
Same text backbone; added vision encoder and aligner; ~305B total. Vision-Exp is all FP8; 0731 can use FP4 experts. Cybergym dips slightly versus the text baseline.
Why not treat the table as a leaderboard?
The harness is unpublished. On some multimodal items the text baseline ignores images. Independent reproductions are not out yet.
How does it sit next to GLM-5.3-Flash?
Both are MIT multimodal MoEs above 300B. GLM weights landed 25 August; DeepSeek ran API first, then weights. Different sequence, not the same checkpoint.
Can sensitive data go to the hosted API?
Usually no for regulated work: requests transit servers in China. Self-hosting the MIT weights removes that outflow; it does not equal an independent security audit of the checkpoint.