DeepSeek V4 Flash Vision Exp: A Vision Experiment, Now Retired

"Exp" in a model name is a warning, and DeepSeek V4 Flash Vision Exp is a clean example of what the warning means. It launched on 21 August 2026. On 10 September DeepSeek retired it on its own API and pointed the alias at a different model. Three weeks is a short life, even by 2026 standards.

That does not make it irrelevant. The weights are still public, some third-party hosts still list it, and the way it was retired is a lesson in what happens to code that calls an experimental model by name.

What it was

DeepSeek described it as the first experimental multimodal model in the V4 family. It takes the V4 Flash architecture, adds visual modules (a vision encoder and an aligner), and continues training so the model can read images alongside text. The base it is compared against is V4 Flash 0731, the July refresh of the text-only model.

DeepSeek's release post pitched it at agent workflows that combine visual understanding with tools, and at multimodal agent tasks, while matching V4 Flash on text work. Text and images can be mixed in one input through the Chat Completions, Messages and Responses formats, with images passed as base64, as a URL or by a Files API id, so adding vision to an existing V4 Flash integration was mostly a model-name change rather than a rewrite.

The weights are on Hugging Face under the MIT licence; the checkpoint metadata shows roughly 305B parameters. DeepSeek does not publish an active-parameter count for the vision model itself, though the V4 Flash backbone it builds on activates about 13B per token. The release post describes it as an API model and says nothing about the DeepSeek app or web chat.

Reading DeepSeek's own table carefully

The launch headline was that the model came "close to Opus-4.8", Anthropic's Claude model, on multimodal agent benchmarks. Summaries of a claim like that tend to drift from "close to" towards "matches", so go to the source. The model card has the actual table, comparing Vision Exp, V4 Flash 0731 and Opus 4.8:

• Terminal Bench 2.1: 83.9, against 82.7 and 85.0
• NL2Repo: 57.7, against 54.2 and 69.7
• Cybergym: 75.3, against 76.7 and 78.3
• DeepSWE: 59.3, against 54.4 and 58.0
• DSBench-Hard: 63.6, against 59.6 and 71.7
• ApexBench: 36.5, against 26.2 and 39.4
• Agents' Last Exam: 27.3, against 25.2 and 25.7
• ZeroBench (pass@5): 35.0, with no base score, against 34.0

Count the rows rather than the adjectives. Across the eleven benchmarks DeepSeek listed, Vision Exp beats Opus 4.8 on three: DeepSWE, Agents' Last Exam and ZeroBench. It trails on the other eight, sometimes narrowly and sometimes, as on NL2Repo, by twelve points. "Close on several, ahead on a few, clearly behind on some" is the honest summary.

Two further details are easy to miss. The base model's ApexBench and Agents' Last Exam scores are marked as runs where it ignored the image content, which is why the gap there looks large: you are partly comparing a model that can see with one that cannot. And Vision Exp scored slightly below its own base on Cybergym, so adding vision was not free on every text task. Vendor-claimed vs independent benchmarks covers how to discount tables like this one.

What "experimental" meant in practice

When DeepSeek released V4.1 Flash on 10 September 2026, it retired both V4 Flash and V4 Flash Vision Exp. Both aliases were set to route temporarily to V4.1 Flash for compatibility. V4.1 Flash has native image input of its own, so for most callers the request still succeeds.

That is the quiet part. A request to deepseek-v4-flash-vision-exp on DeepSeek's API no longer reaches the model this article describes. It reaches a 552B encoder-decoder model with a different architecture, different benchmark profile and different pricing. Nothing errors. Your outputs simply change.

If you integrated it during those three weeks, check three things:

1. Which provider you call. DeepSeek's API reroutes; a third-party host serving the open weights may still run the original model, at least for now.
2. Whether your evaluation baseline is still valid. Any quality numbers you gathered in August describe a model you are no longer getting from DeepSeek directly.
3. Whether you should move to the explicit new ID. Relying on a compatibility alias means a second silent change when the alias is removed.

Pinning model versions and model deprecation and migration cover how to set up integrations so this kind of change is visible.

Is there still a reason to run it

On OpenRouter it is still listed at the time of writing, with a 1M-token context window and around $0.22 per million input tokens and $0.65 per million output. On DeepSeek's own API, before the retirement, the release post said it was billed at V4 Flash pricing, with each image counted as up to 384 tokens.

Reasons to keep using it are narrow. You might self-host the MIT weights because you need a vision-capable model that is smaller than V4.1 Flash's 552B. You might need to reproduce results gathered on it. Or you might be studying how a vision adapter affects a text model's agentic scores, for which a base-plus-vision pair with a published comparison is unusually useful.

For new work, use V4.1 Flash. DeepSeek positions it as the successor with stable vision, and an experimental model whose first-party endpoint has already been retired is not something to build on. Vision-language models compared covers alternatives if you want a vision model outside the DeepSeek line.

How to evaluate any experimental vision model

The lesson generalises. Before committing to an "Exp" or preview model with image input:

• Test on your own images, not the vendor's. Screenshots, scanned PDFs and charts fail in different ways.
• Run the same tasks text-only on the base model to measure what vision costs you on text.
• Keep a fallback model wired in from day one, and log the exact model ID the provider reports back in each response.
• Put a calendar reminder to recheck the provider's changelog monthly.