On 21 August 2026 DeepSeek put a vision checkpoint on its API. Ten days later it put the weights on Hugging Face.
The repo is deepseek-ai/DeepSeek-V4-Flash-Vision-Exp. License is MIT. Parameter count on the card is 305B. This is the first vision model in the V4 Flash family, and it is still marked experimental.
We already covered the API drop. This post is the self-host story: what actually landed, what the vendor table says, and what you should not assume.
What changed on 31 August
The API model and the open checkpoint are the same brain. Visual modules were added on top of V4 Flash, then continued training for screenshots, charts, and UI. Text agent scores stay close to V4 Flash 0731. Multimodal agent scores jump, because the text-only model was guessing without looking.
Hugging Face lists BF16, F32, and quantized shards. Community quants are already appearing. The card is still thin: one evaluative paragraph and a vendor table. No independent eval suite shipped with the weights.
Context window claimed around the release is 1M tokens, with a large max output. Treat that as a serving-stack problem, not a laptop problem. 305B MoE does not fit in a single consumer GPU.
The vendor table, without the slogan
DeepSeek’s own split is the useful part. Text stays flat. Vision is the delta.
Text agent tasks:
| Benchmark | Vision Exp | Flash 0731 | Opus 4.8 |
|---|---|---|---|
| Terminal Bench 2.1 | 83.9 | 82.7 | 85.0 |
| NL2Repo | 57.7 | 54.2 | 69.7 |
| Cybergym | 75.3 | 76.7 | 78.3 |
| DeepSWE | 59.3 | 54.4 | 58.0 |
| Toolathlon-Verified | 75.9 | 70.3 | 76.2 |
| DSBench-Hard | 63.6 | 59.6 | 71.7 |
| AutomationBench (Public) | 25.7 | 25.1 | 27.2 |
Multimodal agent tasks:
| Benchmark | Vision Exp | Flash 0731 | Opus 4.8 |
|---|---|---|---|
| ApexBench Pass@1 | 36.5 | 26.2* | 39.4 |
| Agents’ Last Exam | 27.3 | 25.2* | 25.7 |
| Chartography | 64.3 | - | 65.0 |
| ZeroBench Pass@5 | 35.0 | - | 34.0 |
* Flash 0731 on ApexBench and Agents’ Last Exam ignores the image and guesses from text. Dash means DeepSeek did not publish a Flash number.
Two honest caveats. First, these are DeepSeek’s numbers. Independent writeups already note it trails Opus 4.8 on most of the published rows even while winning a couple of vision-agent tests. Second, NL2Repo is still a hole: 57.7 against 69.7. Vision did not buy you a coding model.
The “exp” tag is doing work. The API string has no 0731-style pin. An experimental build can move between the day you eval and the day you ship.
How you actually send a picture
Three paths, same OpenAI-style content blocks:
- Inline base64
data:URL. Simple. Hits a ~48 MiB request body cap. - Public http(s) URL. DeepSeek fetches it (URL length cap, ~32 MiB image, 60s timeout).
- Files API. Upload once, reuse a file id. Largest images (up to ~64 MiB) and the only sane path if the same screenshot is in a loop.
Images bill as input tokens at V4 Flash rates, capped at 384 tokens per image. There is no vision surcharge. Peak vs off-peak still doubles the dollar card on DeepSeek’s own API. JPEG, PNG, GIF, WebP; format is sniffed from bytes, not from the filename.
FIM completion is not on this checkpoint. Plain Flash still has it in non-thinking mode. If your editor loop depends on FIM, do not silently swap the model id.
What we would run, and what we would not
Self-host this if you already operate MoE serving (vLLM / SGLang class) and you need screenshots inside the same agent that already calls tools. Do not self-host this as a captioning toy. The cost is the 305B footprint and the vision encoder, not the MIT license.
Keep a pinned eval set of your own: one terminal screenshot, one chart, one scanned form, one UI with small type. Vendor ApexBench is not your product.
How IndieRouter fits
IndieRouter is an OpenAI-compatible API on GPUs in India. The interesting property of Vision Exp is not “we beat Opus.” It is one model for text agents plus eyes, billed like Flash, now legally easy to run yourself.
We will not pretend the experimental checkpoint is a GA vision SKU. We also will not pretend you still need a second vision vendor just to read a dashboard. If you are DPDP-bound, the decision is where the screenshot sits, not whether the whale logo is on the card.
Weights: Hugging Face model card. Our earlier API-day note: V4 Flash Vision Exp.