Quick Verdict
DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal version of DeepSeek-V4-Flash-0731, the smaller and cheaper of the two models DeepSeek released back in April 2026 alongside V4-Pro. It went live on the DeepSeek API platform on August 21, 2026, keeping V4-Flash's text performance intact while adding image understanding — reading screenshots, diagrams, and dashboards as part of an agent's workflow — at the exact same per-token price as the text-only model.
This isn't a hands-on usage review — the model has been public for a few days at the time of writing. What follows is based on DeepSeek's own launch benchmarks (published in its official API documentation), plus independent reporting from OfficeChai, The Decoder, and SiliconANGLE. We'll update this page with first-hand testing notes once we've run it against real workloads.
Our early take: 7.8/10 — a genuinely useful, near-free way to add vision to agent workflows already built on V4-Flash, with the caveat that every benchmark cited below comes from DeepSeek testing its own model, and it's explicitly labeled "experimental" rather than a permanent part of the lineup.
What Is DeepSeek V4-Flash-Vision-Exp?
V4-Flash-Vision-Exp is DeepSeek's answer to a gap in its lineup: V4-Pro and V4-Flash, released in April 2026, were text-only. This release doesn't replace either — DeepSeek is explicit that it's "a fairly narrow addition" built for agent use cases that need to see as well as reason, not a new flagship model. It's a multimodal variant of V4-Flash-0731 specifically, built on the same sparse Mixture-of-Experts architecture with 284 billion total parameters and roughly 13 billion active per prompt, which is what keeps inference cheap despite the model's overall scale.
According to DeepSeek, the vision variant holds onto V4-Flash's text capabilities — agentic behavior, reasoning, general knowledge — while adding the ability to process images: describing photos, extracting text from screenshots, and parsing diagrams or dashboards as part of a longer agent task. DeepSeek also shipped version 0.1.1 of its open-source Harness framework the same day, with out-of-the-box support for the new model, and a free Files API that lets developers upload an image once and reuse it across requests by ID instead of resending the payload each time.
Pricing: Same Rate as V4-Flash, Images Billed as Tokens
The pricing pitch is straightforward: DeepSeek didn't create a new price tier for vision. Images are tokenized for billing — capped at 384 tokens each — and charged at the same rate as text tokens on V4-Flash. If you're already building on V4-Flash, adding image input doesn't restructure your cost model.
| Item | Off-Peak | Peak |
|---|---|---|
| Input (cache hit) — per 1M tokens | $0.007 | $0.014 |
| Input (cache miss) — per 1M tokens | $0.22 | $0.44 |
| Output — per 1M tokens | $0.66 | $1.32 |
| Images | Billed as input tokens, capped at 384 tokens/image, same rate as above | |
Peak hours run 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; every other hour is off-peak, at half the peak rate. Context length is 1M tokens with a maximum output of 384K, and the concurrency limit is 2,500 requests — identical to V4-Flash's specs. For comparison, that off-peak cache-miss input rate ($0.22/1M tokens) is a fraction of what Anthropic charges for Opus 4.8, which is the model DeepSeek is benchmarking against.
Try V4-Flash-Vision-Exp on the DeepSeek API platform
Visit DeepSeek Platform →Benchmarks: Where It Beats Opus 4.8, and Where It Doesn't
The numbers below are DeepSeek's own, published in its official API documentation, comparing three models: V4-Flash-Vision-Exp, the text-only V4-Flash-0731, and Claude Opus 4.8. DeepSeek tested using its own "Harness Minimal Mode" with max tokens set to the model's ceiling, top_p at 0.95, and temperature at 1.0 — a setup DeepSeek controls end to end, so treat these as vendor-published figures rather than independently verified results.
| Benchmark | V4-Flash-Vision-Exp | V4-Flash-0731 | Opus 4.8 |
|---|---|---|---|
| Terminal-Bench 2.1 | 83.9 | 82.7 | 85.0 |
| NL2Repo | 57.7 | — | 69.7 |
| ApexBench (Pass@1) | 36.5 | 26.2* | 39.4 |
| Agents' Last Exam | 27.3 | — | 25.7 |
| ZeroBench (Pass@5) | 35.0 | — | 34.0 |
| Chartography | 64.3 | Not tested† | 65.0 |
*V4-Flash's ApexBench score reflects a text-only model ignoring the image inputs in that test. †DeepSeek didn't publish a V4-Flash baseline for Chartography, since the benchmark leans heavily on visual input.
The pattern is consistent: on text-heavy agentic benchmarks like Terminal-Bench 2.1, V4-Flash-Vision-Exp barely moves versus its text-only sibling, and both trail Opus 4.8 by a small margin. On NL2Repo and DSBench-Hard (not shown above, but reported by DeepSeek to trail Opus 4.8 by roughly eight points), the gap to Opus 4.8 widens noticeably. It's specifically on multimodal reasoning — Agents' Last Exam and ZeroBench — that V4-Flash-Vision-Exp actually edges past Opus 4.8, and on ApexBench and Chartography it narrows the gap substantially versus what a text-only model could ever manage on a vision-heavy test.
Key Features
- Vision without a new price tier: images are billed as tokens at V4-Flash's existing rate, so adding multimodal input to an existing agent doesn't change your unit economics.
- 1M-token context, 384K max output: identical to the base V4-Flash model, enough to hold a long agent session or a sizeable document alongside image inputs.
- Built for agent workflows, not chat UI polish: DeepSeek frames this as infrastructure for agents that need to see a dashboard screenshot or parse a diagram mid-task, not a consumer image-chat feature.
- Free Files API: upload an image once and reference it by
file_idacross multiple requests instead of re-sending the same payload every time. - Multiple API formats: supports Chat Completions, Messages, and Responses formats, with mixed text-and-image input via base64, external URLs, or the Files API.
See DeepSeek's official vision API documentation
Read the Vision API Guide →Pros and Cons
| Pros | Cons |
|---|---|
| Adds vision at zero price premium over V4-Flash's already-cheap rate | Labeled "experimental" — not confirmed as a permanent part of DeepSeek's lineup |
| Beats Opus 4.8 on 2 of the 7 published benchmarks (Agents' Last Exam, ZeroBench) | All benchmark figures are vendor-published by DeepSeek, not independently verified |
| Text performance holds steady versus the base V4-Flash model | Trails Opus 4.8 by a wide margin on NL2Repo and DSBench-Hard |
| Free Files API for reusing uploaded images across requests | Only a few days old at time of writing — no independent hands-on testing yet |
| Same 1M-token context and API formats as V4-Flash, easy to adopt if already using it | Images capped at 384 tokens each, which may compress detail on complex screenshots |
Who Should Use DeepSeek V4-Flash-Vision-Exp?
Worth evaluating if you are:
- Already building agents on DeepSeek V4-Flash and want to add screenshot or diagram understanding without restructuring your cost model.
- Cost-sensitive on multimodal API spend — the off-peak cache-miss input rate is a small fraction of what Anthropic or OpenAI charge for comparable vision-capable models.
- Building agent workflows that occasionally need to "see" — reading a dashboard, parsing a chart, confirming a UI state — rather than doing heavy, standalone image analysis.
Probably not the right fit yet if you:
- Need a production-guaranteed model — DeepSeek explicitly labels this "experimental," and whether it graduates into the permanent lineup is still undecided.
- Are doing detail-heavy visual analysis where Opus 4.8's wider benchmark lead (NL2Repo, DSBench-Hard) matters more than price.
- Want independently verified benchmarks before committing — everything published so far comes from DeepSeek's own testing setup.
Frequently Asked Questions
Is DeepSeek V4-Flash-Vision-Exp free?
No, it's pay-per-token through the DeepSeek API, at the same rate as the text-only V4-Flash model: as low as $0.007 per million input tokens off-peak (cache hit) up to $1.32 per million output tokens at peak. Images are billed as input tokens, capped at 384 tokens each. The Files API for reusing uploaded images is free.
Is it actually better than Claude Opus 4.8?
On some benchmarks, yes — DeepSeek's own numbers show V4-Flash-Vision-Exp edging past Opus 4.8 on Agents' Last Exam (27.3 vs 25.7) and ZeroBench Pass@5 (35.0 vs 34.0). On others, like NL2Repo and DSBench-Hard, it trails Opus 4.8 by a wide margin. It's not a broad win, and all figures come from DeepSeek's own testing rather than independent benchmarking.
Does it replace DeepSeek V4-Flash or V4-Pro?
No. DeepSeek describes it as a narrow, experimental addition — a multimodal variant of V4-Flash-0731 — not a new flagship. V4-Pro and V4-Flash remain the primary text-only models in the lineup.
How do I access it?
Through the DeepSeek API platform, using the model identifier deepseek-v4-flash-vision-exp. It supports Chat Completions, Messages, and Responses API formats, with image input via base64, external URLs, or the Files API. DeepSeek's open-source Harness framework (version 0.1.1) added out-of-the-box support the same day the model launched.
How large is the model?
It uses a sparse Mixture-of-Experts architecture with 284 billion total parameters and roughly 13 billion active per prompt — the same scale as the base V4-Flash-0731 model it's built on.
Final Verdict
V4-Flash-Vision-Exp isn't trying to be DeepSeek's next flagship, and it doesn't need to be to matter. It's a low-risk way to extend an already-cheap, already-popular model into agent workflows that need basic image understanding — reading a screenshot, checking a chart, confirming a UI state — without asking developers to pay a multimodal premium or restructure how they're already calling V4-Flash. The benchmark story is genuinely mixed: real wins on two multimodal-reasoning tests against Opus 4.8, a real gap on text-heavy repository and dataset tasks, and every number sourced from DeepSeek's own test harness rather than a neutral third party.
The honest caveat, as with any launch-week coverage: this is based on DeepSeek's published specs and benchmarks, not our own extended testing, because the model has only been public for a few days. We'll revisit this review with hands-on notes once we've run it against real agent workloads.
Early rating: 7.8/10 — a smart, cheap way to add vision to an agent stack already built on DeepSeek, with the normal experimental-release caveats attached.