News / Model Watch · Open-Weights Release
DeepSeek Opens the Vision-Exp Weights: The 1M-Context Vision Agent Goes Self-Hostable, MIT and Ungated
Ten days after the API went live on Aug 21, DeepSeek posted the weights: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp landed on Hugging Face on Aug 31 — MIT licensed, not gated, with the model card's benchmarks matching the API changelog number for number. The V4 family's experimental multimodal agent is now yours to download.
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp weights are live on Hugging Face (Aug 31): MIT license, ungated, 84 files with 49 safetensors shards, plus the repo's encoding toolkit and eval configs (Terminal-Bench 2.1, Deep-SWE). Config: 43-layer MoE, 256 routed experts with 6 active per token, 1,048,576-token context — same 1M window as the API that launched Aug 21. Card benchmarks match the API changelog exactly (TB 2.1 83.9).
What's in the repo (API-verified)
- The weights. 49 safetensors shards across 84 files; not gated, no acceptance queue — clone and go. Upload began ~14:16 Beijing on Aug 31 and files were still landing as of ~20:23.
- The license. MIT, plain and simple. Compare GLM-5.3's custom license or the "other" tags on Qwen's Flash-Next repos — DeepSeek keeps shipping the cleanest terms on the shelf. Commercial self-hosting: no extra reading required.
- The architecture. config.json: 43 hidden layers, 256 routed experts with 6 active per token (MoE), hidden size 4096, vocab 129,280, and max_position_embeddings 1,048,576 — the full 1M-token window, matching the V4 line's context story.
- The receipts. The repo ships eval_results configs for Terminal-Bench 2.1 and Deep-SWE — the same benchmarks on the Aug-21 API changelog entry. Card numbers match that changelog exactly: this is the same model developers have been calling for ten days.
- The tokenizer. A custom
encoding/toolkit (encoding_dsv4.py with tests) rides along — useful for anyone doing local serving or research on the byte-level side.
Vision-Exp vs its own base vs Opus 4.8 (official card)
| Benchmark | Vision-Exp | V4-Flash-0731 | Opus 4.8 |
|---|---|---|---|
| Terminal Bench 2.1 | 83.9 | 82.7 | 85.0 |
| NL2Repo | 57.7 | 54.2 | 69.7 |
| DeepSWE | 59.3 | 54.4 | 58.0 |
| DSBench-Hard | 63.6 | 59.6 | 71.7 |
| ApexBench (multimodal, Pass@1) | 36.5 | 26.2 | 39.4 |
| Agents' Last Exam | 27.3 | 25.2 | 25.7 |
| Chartography | 64.3 | — | 65.0 |
Vendor-reported numbers from the official model card (Aug 31). Pattern worth noting: the vision modules add roughly +4 points on multimodal agent tasks over the text-only Flash base while keeping text-agent performance within ~1 point — and DeepSWE actually clears Opus 4.8.
What it completes on the open-weight shelf
With tonight's drop, every model in DeepSeek's live API lineup now has open weights: V4-Flash (Jul 31), V4-Pro-0813 (Aug 13), and the Vision-Exp experimental multimodal (Aug 31) — all MIT. No other major Chinese lab closes the API-to-weights loop this fast across its whole catalog; GLM-5.3's weights carry a custom license, and Qwen's Flash-Next preview still awaits its official announcement. For builders, the practical read: you can call any DeepSeek model via API today and self-host the same weights tomorrow — with Vision-Exp being the first 1M-context vision agent in that set, and community quantizations presumably days away.
Primary sources
FAQ (2026)
Weights really out?
Yes — uploaded Aug 31 (~14:16 Beijing), ungated, 49 safetensors shards + encoding toolkit + eval configs. Still updating as of ~20:23.
License?
MIT — no custom clauses, commercial self-hosting fine. Cleanest terms among current Chinese frontier weights (GLM-5.3 uses a custom license).
How big? Self-hostable?
43-layer MoE, 256 experts / 6 active, 1M context. Full-precision is multi-hundred-GB (49 shards); community quantizations likely within days.
Same as the API model?
Yes — card benchmarks match the Aug-21 changelog exactly (TB 2.1 83.9, DeepSWE 59.3). API: model='deepseek-v4-flash-vision-exp'.
vs Opus 4.8?
Per DeepSeek's card: text-agent close (83.9 vs 85.0 TB2.1), DeepSWE ahead (59.3 vs 58.0), multimodal agent tasks up big over its own base (+10.3 ApexBench). Directional — vendor numbers.
Why it matters?
Every live DeepSeek API model now has MIT weights (Flash 07-31, Pro 08-13, Vision-Exp 08-31). First 1M-context vision agent on the open shelf.