DeepSeek-V4-Flash-Vision-Exp Goes Open Weights: Same Price as V4-Flash, Now Self-Hostable

DeepSeek-V4-Flash-Vision-Exp, the experimental multimodal model that launched on DeepSeek's API on August 21, released its weights on Hugging Face on August 31 under the MIT license. Ten days after the API launch, the same brain that bills at exactly the token rates of DeepSeek-V4-Flash is downloadable in full — model files, tokenizer, prompt-encoding reference, a minimal PyTorch inference path, and day-0 vLLM serving, which the vLLM project announced the same day. That makes the cheapest 1M-context agent with eyes also the cheapest one to self-host, and it means the reference implementation is no longer the only way to run it. The one new claim worth flagging up front: analyst Teortaxes, reading the release on X, said the open-weights release brings feature parity with the vision models from Moonshot and GLM. The open weights are verified fact; that parity is a third-party reading nobody has reproduced yet.

What actually shipped

DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model, accessed with the model ID deepseek-v4-flash-vision-exp. Architecturally it is a vision encoder and an aligner on the V4-Flash MoE backbone — the same roughly 284-billion-parameter, 13-billion-active text model, with visual modules added. It takes text plus images in and produces text out, over a 1M-token context window with a 384K maximum output, in both thinking and non-thinking modes. It supports the Chat Completions format, Anthropic-style Messages, and the Responses format, which is the trio an agent framework needs to wire it in without a custom adapter.

On image input, DeepSeek accepts JPEG, PNG, GIF, and WebP, passed three ways: base64 inline, an external URL, or a file uploaded through the Files API. There is no video or audio input yet — the "vision" here is stills only.

DeepSeek's own positioning, in the release note: pure-text capability — agents, reasoning, world knowledge — is on par with the official DeepSeek-V4-Flash release, while multimodal agent capability takes a major leap and lands close to Opus-4.8. That is a vendor claim, unreproduced, and worth reading carefully, because the benchmark detail behind it is more interesting than the headline.

What the open-weights release actually includes

Today's release answers the question the launch post left open. The Hugging Face repository — deepseek-ai/DeepSeek-V4-Flash-Vision-Exp — carries the MIT license and ships the full weights (roughly 285 billion parameters — the V4-Flash MoE backbone plus a vision encoder and aligner — about 13 billion active per token, split across 48 shards), the tokenizer, an OpenAI-style prompt-encoding reference implementation, and a minimal PyTorch inference path covering the vision encoder, the aligner, DFlash attention, the MoE, Hyper-Connections, and the DSpark forward pass. DeepSeek still calls the PyTorch code a reference implementation rather than a production serving engine — but it no longer has to be the serving path, because vLLM announced day-0 support the same day.

The vLLM project's own statement is that it serves the model now. What that means in practice, verified against the vLLM repository and the model card, is a pinned day-0 docker image (vllm/vllm-openai:deepseekv4-flash-vision) and an official recipe tied to a still-open upstream pull request (#54566). Vision support is not in a released vLLM wheel — the plain wheel routes the checkpoint to the text-only path and fails on the vision tensors — so day-0 here means an image you pin, not an upgrade to an existing vLLM install. On that image, vLLM reports a working vision path on NVIDIA GB200 and an OCRBench score of 835 (83.5%) on its reference configuration. The practical ceiling has moved from "wait for a stable release" to "download, serve, and measure it yourself."

On X, analyst Teortaxes framed the release in competitive terms: the open weights now give DeepSeek some feature parity with the vision models from Moonshot and GLM. That is a third-party reading — not a DeepSeek statement and not a measured result — but it is testable in a way it was not before this release, because the same release hands anyone the files to run the comparison.

Why the benchmark jump is bigger than it looks

All of the following figures are DeepSeek's own, published in the launch note, not independently verified. On the agent benchmarks that embed screenshots and images, the text-only DeepSeek-V4-Flash does not read them — it simply ignores the multimodal parts of the task. DeepSeek-V4-Flash-Vision-Exp actually looks, which is why the deltas are so large:

• ApexBench Pass@1 — Vision-Exp 36.5 vs V4-Flash 26.2 (Opus-4.8 39.4)

• Agents' Last Exam — Vision-Exp 27.3 vs V4-Flash 25.2 (Opus-4.8 25.7)

• Terminal-Bench 2.1 — Vision-Exp 83.9 vs V4-Flash 82.7 (Opus-4.8 85.0)

• NL2Repo — Vision-Exp 57.7 vs V4-Flash 54.2 (Opus-4.8 69.7)

• DeepSWE — Vision-Exp 59.3 vs V4-Flash 54.4 (Opus-4.8 58.0)

• Chartography — Vision-Exp 64.3 (text-only V4-Flash cannot attempt it)

• ZeroBench Pass@5 — Vision-Exp 35.0 (text-only V4-Flash cannot attempt it)

Two of those numbers deserve a second look. On Agents' Last Exam, Vision-Exp (27.3) edges Opus-4.8 (25.7), and on DeepSWE it beats Opus-4.8 (59.3 vs 58.0) as well. That is what "close to Opus-4.8" actually rests on — not one marquee result, but a band of agentic benchmarks where seeing the screen closes most of the gap to the frontier multimodal model while underselling it on price by roughly an order of magnitude.

What an image costs, to the decimal

Images are tokenized for billing — up to 384 tokens each — and then charged at the same per-token rates as DeepSeek-V4-Flash. Because DeepSeek moved all of its models to peak/off-peak pricing on August 16, 2026, the exact numbers depend on when you call:

• Input, cache miss — $0.22 per 1M tokens off-peak, $0.44 peak

• Input, cache hit — $0.007 per 1M tokens off-peak, $0.014 peak

• Output — $0.66 per 1M tokens off-peak, $1.32 peak

• Peak hours — 01:00–04:00 and 06:00–10:00 UTC (all other hours off-peak)

Do the arithmetic and the cost of vision nearly vanishes. One image at its 384-token cap is about 0.0085 cents off-peak and 0.017 cents at peak, as input. An agent that reads 50 screenshots in a single run is adding roughly half a cent of input tokens — before the model even writes its first output token. DeepSeek also clamps resolution before inference: the low detail level resizes to 512×512, and high/original pre-scales the image (small ones up to about 384×384, large ones down to about 800×800), so a high-resolution original is never fed to the model at its native pixel count.

Those rates are now a ceiling rather than the only option. With the weights public, self-hosting costs whatever your GPUs cost instead of the per-token list price — worth running the numbers on if you already operate inference hardware, irrelevant if you do not.

The supporting cast: a free Files API and Harness 0.1.1

Two infrastructure pieces shipped alongside the API launch. The Files API is live and free: upload an image once, reference it by file_id, and reuse it across requests without re-sending the bytes — meaningful for a browsing agent that holds a page in context across several turns. And DeepSeek Harness 0.1.1, released the same day, has out-of-the-box support for the new model, so the harness that benchmarks and drives DeepSeek agentic workloads can target it immediately. With the weights MIT-licensed, the model-plus-toolchain stack is now open end to end.

The production caveat, stated plainly

This is an experimental model. DeepSeek labels it explicitly as such, has announced no GA date, and recommends against running it directly in production; the stated plan is to iterate on feedback and ship a stable version later. The practical consequence is that the text brain is proven — it is the same weights family that powers V4-Flash — but the vision path is new code. vLLM's day-0 image makes it runnable today, yet that image is explicitly pre-release: the vision implementation lives in an open pull request, not a stable release, and the only hardware with a published vision run so far is NVIDIA GB200. So the first wave of independent stress-testing is still just beginning this week — but it is beginning with a real serving engine, not just DeepSeek's reference code.

That is exactly the situation where a routing layer earns its keep. On OrcaRouter, both deepseek/deepseek-v4-flash-vision-exp and deepseek/deepseek-v4-flash are live behind one API, at DeepSeek's list price passed through with zero markup. You can point image-bearing traffic at the experimental model and, in the same configuration, set an automatic failover to a fallback the moment it errors or times out — which de-risks the whole "new code" problem. The routing DSL also lets you send only the calls that actually contain images to the vision model and keep pure-text traffic on DeepSeek-V4-Flash, capturing the benchmark gains where they exist and paying nothing extra where they don't.

The MIT license also means the self-host path now exists for teams that want it — the release includes reference inference code, and vLLM's day-0 image adds a production-grade serving option, so you are not locked into the API. For everyone who does not want to operate a roughly 285-billion-parameter MoE just to evaluate a model, the routing API is the lower-friction on-ramp.

What to watch next

The open questions have narrowed since the API launch. Does DeepSeek-V4-Flash-Vision-Exp reach GA, and does the price hold? Does video or audio input follow — the model is stills-only today, which leaves the large multimodal frontier models unopposed on moving media? When does vLLM's vision path graduate from the pinned day-0 image into a stable release, and onto hardware beyond GB200? And do the independent evaluations — the first third-party run on ApexBench or Terminal-Bench, now possible for anyone with GPUs — reproduce the published numbers?

One thing has changed since the release note went out. On September 8, Teortaxes — the analyst whose parity framing appears above — reported from hands-on testing that DeepSeek now appears to run at least two modern V4-Flash-Vision models. The one he reached through gray testing was, in his testing, the stronger of the two, and it ran at the same speed as DeepSeek-V4-Flash-Vision-Exp, which he read as a shared base; a newer build is faster but weaker. That is a single-researcher gray-test observation, not a DeepSeek statement and not an independent benchmark. It landed the same day DeepSeek opened a two-day API beta of an “intermediate version” it bills as native-multimodal and faster — the model ID deepseek-v4.1-flash-expires-on-0910 began serving on September 8 — so the practical question for a buyer is where the release build lands on the speed-versus-strength trade: whether the faster tier keeps the agentic benchmark profile that made the open-weights release interesting, or trades it for throughput.

The Moonshot/GLM parity question is the other one to watch. Teortaxes' framing is the testable version of the vendor's "close to Opus-4.8" line: with the weights open, a direct comparison against Moonshot's and GLM's vision models can now be run by anyone on their own benchmarks. Until someone does, it stays a claim. What does not need reproducing is the pricing fact underneath the whole release — vision at text prices is a billing number, not a benchmark, and with the weights open it now has a free alternative attached.