Vision-Exp is the pick for teams already on DeepSeek-V4-Flash who want cheap document and chart understanding bolted onto an existing agentic pipeline, not a flagship vision model. It trails Opus 4.8 on most benchmarks DeepSeek disclosed at launch, including a 12-point gap on NL2Repo, and ships with no open weights.
DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model DeepSeek released August 21, 2026, adding image understanding to its 284B-parameter, 13B-active MoE V4-Flash. It holds a 1,048,576-token context window and beats Claude Opus 4.8 on 3 of 11 disclosed multimodal benchmarks, including ZeroBench Pass@5 (35.0 vs 34.0).
Provider: DeepSeek · Family: DeepSeek-V4
Context window: 1,048,576 tokens · Max output: 384,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About DeepSeek-V4-Flash-Vision-Exp
DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model that DeepSeek, the Hangzhou-based AI lab founded by Liang Wenfeng in 2023, released on August 21, 2026 through its API platform. It is a vision-enabled fork of DeepSeek-V4-Flash-0731, the fast, cost-optimized sibling in the DeepSeek-V4 line (284 billion total parameters, 13 billion active per token, Mixture-of-Experts with Hybrid Attention combining Compressed Sparse Attention and Heavily Compressed Attention). Rather than a new flagship, DeepSeek built it to add image understanding onto V4-Flash's existing agentic and reasoning stack without changing the underlying text model. On text-only evaluation, Vision-Exp matches the base V4-Flash: 88.7% on MMLU in standard mode, with DeepSeek's own technical report additionally listing 86.2% MMLU-Pro, 88.1% GPQA Diamond, and 79.0% SWE-bench Verified in its higher-effort reasoning mode, though those three figures have not been independently reproduced outside DeepSeek yet. On the multimodal and agentic comparison table DeepSeek published alongside the launch, benchmarked directly against Claude Opus 4.8, Vision-Exp trails on most of the 11 disclosed evaluations: Terminal Bench 2.1 sits at 83.9 against Opus 4.8's 85.0, NL2Repo trails at 57.7 versus 69.7, DSBench-Hard is down noticeably too, ApexBench Pass@1 is 36.5 versus 39.4, and Chartography is 64.3 versus 65.0. It edges ahead on three narrower evaluations: Agents' Last Exam, DeepSWE, and ZeroBench Pass@5. DeepSeek itself flags these as self-reported numbers from a controlled test environment, not independently verified by a third party. The context window is 1,048,576 tokens with a 384,000-token max output, identical to text-only V4-Flash. Every image is tokenized at a flat per-image count no matter its resolution, so a request can carry hundreds of images and long text inside the same context budget without a separate vision allowance. Input accepts text and static images via base64, an external URL, or the newly launched Files API; output stays text-only. A single request can hold hundreds of images, with a higher pixel-per-side ceiling when fewer images are attached and a lower one for large batches. There is no native audio, video, or PDF input, so documents need to be rasterized to images first. The model supports Chat Completions, Messages, and Responses API formats, carries over V4-Flash's function calling and JSON mode, and runs with thinking (reasoning) mode enabled by default. DeepSeek added vision to V4-Flash without introducing a separate price tier: input and output rates for Vision-Exp match text-only V4-Flash exactly, and images ride on the same per-token pricing as any other input. Rates follow a two-clock schedule that doubles during two short daily windows and reverts to a lower rate the rest of the time; the exact figures are in the pricing FAQ below. The model is live on DeepSeek's first-party API under the identifier deepseek-v4-flash-vision-exp, and was listed the same day on OpenRouter and tagged Experimental on Vercel's AI Gateway. There is no confirmed listing yet on AWS Bedrock, Google Vertex AI, or Azure. Unlike the MIT-licensed, Hugging-Face-hosted V4-Flash it extends, Vision-Exp itself ships no open weights or public model card, so access is API-only under DeepSeek's standard platform terms. DeepSeek has not published a dedicated system card for this variant at launch. Independent safety research on the DeepSeek model family has generally found lighter default refusal behavior and constitutional alignment than Western frontier labs, and DeepSeek's own retention policy keeps API inputs for a month for abuse monitoring unless a customer is on a zero-retention enterprise plan; the company has not disclosed SOC 2, GDPR, or HIPAA certification for its API platform. Vision-Exp fits teams already running V4-Flash agent pipelines who want document, chart, or screenshot understanding added without a new vendor or a new price tier. It is a weaker choice for anyone needing flagship-tier visual reasoning, since it loses on most of the benchmarks DeepSeek disclosed against Opus 4.8, for production systems that need stability guarantees given DeepSeek's own recommendation to keep a fallback model ready, or for teams that require self-hosted or air-gapped deployment, since this fork has no published weights.
Pricing
Off-peak: $0.22/M input (cache miss), $0.007/M (cache hit), $0.66/M output, identical to text-only V4-Flash with no vision surcharge. A surge schedule doubles those rates twice on weekday mornings UTC; weekends and every other hour stay at the base rate. Each image counts toward the input total at a flat 384-token rate no matter how large the file is.
Key Features
- Million-Token Context With Vision: Accepts text and images together up to the model's full context window, letting one request carry hundreds of images alongside a long document without truncation.
- Vision At No Extra Cost: Image understanding bills at the same off-peak rates as text-only V4-Flash, with no separate vision surcharge.
- Files API For Repeat Images: Upload an image once via the free Files API and reference it by file_id across multiple requests instead of resending raw image data every call.
- Agentic Visual Reasoning: Matches DeepSeek-V4-Flash on agentic, coding, and reasoning tasks while adding document and chart understanding, with multimodal agent benchmarks approaching Opus-4.8 on some evals.
- Large Batch Image Requests: Supports hundreds of images in a single request, with the allowed resolution per image scaling down as the batch of attached images grows larger.
Pros
- Adds image understanding to V4-Flash at identical per-token pricing, with no separate vision surcharge.
- Full text capability of V4-Flash carries over unchanged, including its agentic tool-use behavior.
- Comes out ahead of Claude Opus 4.8 on a few of DeepSeek's disclosed multimodal benchmarks, despite costing dramatically less per token.
- Free Files API lets teams cache and reuse uploaded images across a session instead of paying to re-transmit raw image payloads.
Cons
- Trails Claude Opus 4.8 on most disclosed benchmarks, most notably NL2Repo and DSBench-Hard, where the accuracy gap is largest.
- No open weights or Hugging Face model card for this variant, unlike its MIT-licensed V4-Flash sibling, so self-hosting isn't possible.
- Output remains text-only with no image generation, and there's no native video, audio, or PDF support, only static images.
Benchmarks
- mmlu: 88.7
- nl2repo: 57.7
- mmlu pro: 86.2
- humaneval: 69.5
- chartography: 64.3
- gpqa diamond: 88.1
- agents last exam: 27.3
- swe bench verified: 79
- terminal bench 2 1: 83.9
- apexbench pass at 1: 36.5
- zerobench pass at 5: 35
- dsbench hard gap points: -8.1
- artificial analysis speed tokens per sec: 88
Frequently Asked Questions
What does DeepSeek-V4-Flash-Vision-Exp actually cost?
Off-peak pricing is $0.22 per 1M input tokens on a cache miss, $0.007 per 1M on a cache hit, and $0.66 per 1M output tokens, exactly matching text-only DeepSeek-V4-Flash with no vision surcharge. Rates double automatically during two daily windows (01:00-04:00 and 06:00-10:00 UTC, Monday through Friday); every other hour, including weekends, stays at the off-peak rate. Images bill as input tokens at a flat per-image rate regardless of resolution.
Can you use DeepSeek-V4-Flash-Vision-Exp without paying?
No dedicated free tier exists for this API model; every request is billed at the per-token rate from the first call. DeepSeek's separate consumer chat web app offers free access to some of its models, but this experimental endpoint is API-only and isn't exposed there. Testing it requires a funded DeepSeek Platform API key or an aggregator like OpenRouter.
What are DeepSeek-V4-Flash-Vision-Exp's closest competitors?
DeepSeek-V4 Flash itself is the closest sibling and costs the same per token if a request never needs images. GPT-5 and Claude Opus 4.8 are the flagship-tier vision alternatives when accuracy matters more than price, both ahead on most of the benchmarks DeepSeek disclosed at launch. Neither Vision-Exp nor its base model can be self-hosted for vision workloads, since no open weights exist for this fork.
How does DeepSeek-V4-Flash-Vision-Exp compare to Claude Opus 4.8 in 2026?
DeepSeek's own launch comparison puts Vision-Exp behind Opus 4.8 on most disclosed benchmarks, with the largest gaps on NL2Repo (57.7 vs 69.7) and DSBench-Hard. On the three where DeepSeek reports a lead, Agents' Last Exam, DeepSWE, and ZeroBench Pass@5, the margins are narrow. The trade-off is price: Vision-Exp runs at a small fraction of Opus 4.8's per-token cost for teams that don't need the accuracy gap closed.
How do you set up DeepSeek-V4-Flash-Vision-Exp?
Create a DeepSeek Platform account, generate an API key, and call the Chat Completions, Messages, or Responses endpoint with the model set to deepseek-v4-flash-vision-exp. Images can be sent as base64, an external URL, or uploaded once to the free Files API and referenced by file_id across later turns. No separate vision endpoint or extra authorization is needed since it uses the same key and base URL as text-only V4-Flash.
Top Alternatives
- Claude Opus 4.8: Pick Claude Opus 4.8 if you need the accuracy lead on most disclosed multimodal benchmarks; pick Vision-Exp if price per token matters more than closing that gap.
- GPT-5: Pick GPT-5 for broader multimodal support (audio, wider ecosystem); pick Vision-Exp when you're already spending on V4-Flash and want vision at the same rate.
- DeepSeek-V4 Flash: Pick text-only V4-Flash if you never need images and want the MIT-licensed, self-hostable version; pick Vision-Exp only when a request actually includes an image.