DeepSeek-V4-Flash-Vision-Exp

What it is

DeepSeek-V4-Flash-Vision-Exp is a model that reads images as well as text, added to DeepSeek’s API on 21 August 2026. The company’s changelog announces it in two sentences: “Today, the new multimodal vision understanding model DeepSeek-V4-Flash-Vision-Exp is now available on the DeepSeek API platform. This is an experimental model that can be accessed by setting model='deepseek-v4-flash-vision-exp'.”

It is a sibling of DeepSeek-V4-Flash, the cheaper of the two tiers in the DeepSeek V4 line, rather than a new foundation. It inherits the tier’s specification: a one-million-token context window, output up to 384,000 tokens, and the same price to the cent. What it adds is an eye. Images are turned into tokens according to their dimensions and billed as input.

DeepSeek publishes no parameter count for this variant. The text Flash model it is built beside runs 284 billion total parameters with 13 billion active, and the company has not said whether those figures carry over.

What DeepSeek claims for it

The pitch in the changelog is about agent work, and it is worth reading in full:

“In terms of pure-text capabilities (agent, reasoning, world knowledge, etc.), DeepSeek-V4-Flash-Vision-Exp is on par with the official DeepSeek-V4-Flash. On agent benchmarks that require visual understanding, DeepSeek-V4-Flash-Vision-Exp delivers a significant leap over DeepSeek-V4-Flash, bringing its multimodal agent capabilities close to Opus-4.8.”

DeepSeek changelog, 21 August 2026

The published scores, all of them DeepSeek’s own runs, are these: Terminal Bench 2.1 at 83.9, NL2Repo at 57.7, DeepSWE at 59.3, DSBench-Hard at 63.6, AutomationBench on its public set at 25.7, ApexBench at 36.5 pass@1, Agents’ Last Exam at 27.3, Chartography at 64.3 and ZeroBench at 35.0 pass@5.

DeepSeek prints its own method alongside them, which is more than most launch tables do. Its footnote: “For the Code Agent text tasks in the public benchmark sets, the DeepSeek family models were tested using the DeepSeek Harness minimal mode as the framework, with the max effort level, topp=0.95, and temperature=1.0; in the ApexBench and Agents’ Last Exam evaluations, the text model DeepSeek-V4-Flash ignores the multimodal elements within them.” So the comparison against the text Flash model on those two rows is a comparison against a model that could not see the pictures, which is the point being demonstrated and also a reason the margin looks as wide as it does.

A rival lab reproduced the figures

Launch tables usually stand alone until an independent evaluator gets an endpoint. This one did not have to wait. When Z.ai published GLM-5.3-Flash, the smaller sibling of GLM-5.3, on 26 August 2026, its comparison table included this model, listed as DeepSeek-V4-Vision-Exp, and five of the rows overlap with DeepSeek’s own: Terminal Bench 2.1 at 83.9, NL2Repo at 57.7, DeepSWE at 59.3, Agents’ Last Exam at 27.3 and Chartography at 64.3.

Every one of the five is identical to the figure DeepSeek published. A competitor running a model on its own harness and arriving at the same numbers as the model’s maker is a stronger signal than either table on its own, and it is unusual enough to record.

No weights, which is the departure

DeepSeek built its reputation on publishing the thing itself. The V4 preview was open-sourced on the day it launched in April 2026, and DeepSeek V4 Pro reached general availability in August under an MIT licence with all 66 checkpoint shards on Hugging Face.

Nothing has been published for this model. As of 30 August 2026, the newest repositories in the deepseek-ai organisation on Hugging Face are DeepSeek-V4-Pro-0813, uploaded on 13 August 2026, and DeepSeek-V4-Flash-0731 from 31 July. There is no Vision-Exp repository under any uploader. The model exists only as an endpoint, which is consistent with the experimental label and still a change of habit for this lab.

The API surface is narrower than the text models too. JSON output, tool calls, the OpenAI Responses format, the Anthropic format and chat prefix completion all work. Fill-in-the-middle completion, which the text Flash and Pro models support, does not. The concurrency limit is 2500.

What it costs

The price is DeepSeek-V4-Flash’s price, unchanged, and it runs on the peak and off-peak clock the whole V4 family moved to at 16:00 UTC on 16 August 2026.

Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Everything outside those seven weekday hours bills at the off-peak rate, weekends included, so a batch that can wait until midday UTC costs half what the same batch costs at eight in the morning. Against the flagship, this tier is a third of the price: V4 Pro bills $0.66 and $1.32 for input on a cache miss, and $1.98 and $3.96 for output, on the same clock.

What to watch

Whether the “Exp” suffix comes off. An experimental endpoint can be withdrawn without ceremony, and building a product on one is a bet. The two things that would settle its status are a stable, unsuffixed vision model in the price table, and a weights release: if DeepSeek publishes this one the way it published everything else in the V4 line, the API-only period reads as a staging step. If it does not, the lab’s open-weight default has an exception in it.

For the tier this variant sits inside and the mixture-of-experts design underneath it, see DeepSeek V4; for the flagship on the same billing clock, DeepSeek V4 Pro.