On August 21, DeepSeek began offering an experimental model called "DeepSeek-V4-Flash-Vision-Exp" through its API, adding image input to the previously text-only V4-Flash line. Each image input is capped at a maximum of 384 tokens, and by default, larger images are normalized to roughly 800×800 pixels. Agents that handle visual information—reading screens, charts, or forms—have often struggled with unpredictable input volume and cost. This design places a ceiling right at that entry point. However, true to its name, this remains an experimental release: no weights, architecture details, or technical report are publicly available. Small margins on a benchmark table are not proof that the model can accurately read fine visual detail.
A Vision-Enabled API With Undisclosed Model Details
DeepSeek-V4-Flash-Vision-Exp is described as adding visual input capability to V4-Flash's existing text abilities. It's accessible through the OpenAI-compatible Chat Completions/Responses API and the Anthropic-compatible Messages API. In Chat Completions, images are placed within a user message. In the Responses API, images can also be passed through the user, developer, or tool-output channels. Specifying an unsupported role or model returns an HTTP 400 error, meaning existing conversation histories can't necessarily be swapped in as-is.
Supported formats include JPEG and PNG, as well as GIF and WebP—format is determined by actual file content rather than file extension or declared MIME type. Images can be submitted three ways: inline Base64, a public URL, or a file_id from the Files API. Public URLs are limited to 8,192 characters, images must be 32MiB or smaller, and downloads time out after 60 seconds. Through the Files API, images up to 64MiB are accepted. When connecting to systems that handle images, implementations need to account for where data is retrieved and which failures should trigger retries.
On the same day, DeepSeek also released a free Files API and DeepSeek Harness 0.1.1, which includes built-in support for the new model. The Files API allows storage of up to 25GiB and 10,000 files per user, with an expiration setting configurable from 3,600 to 2,592,000 seconds (one hour to 30 days) at upload time. Omitting this setting results in permanent storage. That said, only the Files API itself is free—inference tokens are billed separately. Reusing a file_id saves upload bandwidth, but the documentation doesn't indicate whether it also reduces prompt-caching costs.
Public documentation does not disclose Vision Exp's parameter count, training methodology, or model architecture. No open-weight link has been provided either. The fact that the API is now available is a separate matter from whether vision capability was simply bolted onto the existing V4-Flash-0731 weights, or whether there's any guarantee of stable production performance.
Three Wins Out of 11 Metrics, a 2-2 Split on the Four Visual Metrics
DeepSeek's published chart compares Vision Exp, V4-Flash-0731, and Anthropic's Opus 4.8 across 11 rows. Vision Exp outperformed Opus 4.8 on three: DeepSWE (59.3 vs. 58.0), Agents' Last Exam (27.3 vs. 25.7), and ZeroBench Pass@5 (35.0 vs. 34.0)—margins of 1.3, 1.6, and 1.0 points, respectively. Opus 4.8 led on the remaining eight rows, with the largest gaps on NL2Repo (57.7 vs. 69.7, a 12.0-point difference) and DSBench-Hard (63.6 vs. 71.7, an 8.1-point difference). Nine of the 11 rows show gaps of 3.0 points or less, but this doesn't amount to evidence of overall parity.
Looking only at the four rows involving images: on ApexBench Pass@1, Vision Exp trails by 2.9 points (36.5 vs. 39.4). On Agents' Last Exam, it leads by 1.6 points. On Chartography, it trails by 0.7 points (64.3 vs. 65.0). On ZeroBench Pass@5, it leads by 1.0 point. So on vision-inclusive evaluations, the record is 2-2. Chartography tests visual recognition and reasoning using 100 charts spanning multiple specialized fields. Agents' Last Exam targets long-horizon, economically valuable real-world tasks with verifiable outcomes. These two benchmarks measure fundamentally different kinds of work.
| Metric | Vision Exp | Opus 4.8 | Gap vs. Opus 4.8 |
|---|---|---|---|
| ApexBench Pass@1 | 36.5 | 39.4 | -2.9 |
| Agents' Last Exam | 27.3 | 25.7 | +1.6 |
| Chartography | 64.3 | 65.0 | -0.7 |
| ZeroBench Pass@5 | 35.0 | 34.0 | +1.0 |
Across the nine metrics shared with V4-Flash-0731, Vision Exp improved on eight of them; only Cybergym dropped, falling from 76.7 to 75.3, a decline of 1.4 points. The largest apparent gain was on ApexBench, at 10.3 points—but DeepSeek's own footnote states that the baseline text model ignores the multimodal components of ApexBench and Agents' Last Exam. This means the gap cannot be attributed solely to the effect of the vision module.
It's also worth noting these figures come from DeepSeek itself. The public coding-agent tasks in the DeepSeek series were run using Harness's minimal mode, a maximum token setting, top_p=0.95, and temperature=1.0. These numbers reflect the combined effect of the model, the harness, and the tool configuration—not the model in isolation. The 35.0 figure on ZeroBench is a pass@5 score with tool use, sourced from vendor reporting rather than an official evaluation run by the benchmark's own site. Full execution logs, sample sizes, and confidence intervals have not been disclosed. Latency and throughput are also unknown.
What Does a 384-Token Image Input Actually Cost?
Vision Exp's pricing and concurrency limits match those of the current V4-Flash. Context length is 1 million tokens, output is capped at 384,000 tokens, and concurrency is set at 2,500. Input pricing depends on whether it's a cache hit or miss. Off-peak, the rate per million tokens is $0.007 for cache hits, $0.22 for cache misses, and $0.66 for output. During peak hours, these rates rise to $0.014, $0.44, and $1.32, respectively. Peak hours are UTC 01:00–04:00 and 06:00–10:00.
Calculating for an image that uses the full 384-token allotment as a cache-miss input: a single image costs $0.00008448 off-peak and $0.00016896 at peak. For a batch of 600 images—230,400 tokens of image data alone—the same assumptions yield $0.050688 off-peak and $0.101376 at peak. Neither figure includes text input or output costs, and in practice, images must also satisfy byte-size and dimension limits.
This cap is also a deliberate design choice to prevent images from expanding into unbounded token counts. Selecting detail=low shrinks the image to 512×512. With the default or "original" setting, larger images are normalized to roughly 800×800 pixels while preserving aspect ratio. This makes input costs easier to control, but there's no guarantee that fine text, complex tables, or dense chart detail survives the process. The 384-token billing cap is not proof of OCR accuracy, chart-reading accuracy, or response speed.
The 600-Image Cap Creates Implementation Trade-offs
A single request can include up to 600 images, but each image's longest side is capped at 8,192 pixels. For requests containing 15 or more images, that cap drops to 4,096 pixels. The total size of images not using file_id is capped at 64MiB; including a file_id raises that to 200MiB. Inline request bodies using Base64 or similar are capped at 48MiB. Building a batch based solely on the 600-image count risks hitting a failure first—on byte size, request body size, or resolution.
Given this, a reasonable split emerges: URL or Base64 for processes that pass along a small number of screenshots or diagrams immediately, and the Files API for processes handling large numbers of reused images. Still, public URLs become unreadable if the source takes longer than 60 seconds to respond. Base64 eats into the body-size limit. And the Files API requires managing both storage caps and expiration settings. Each input pathway shifts not the cost, but the type of failure you're likely to encounter.
Vision Exp does not support FIM. The model table lists thinking and non-thinking modes, JSON output, and tool calling. It also supports the Responses API and Anthropic-compatible interfaces, but that doesn't guarantee the same functionality can simply be added to an existing code-completion pipeline. Anyone evaluating an agent that factors in images needs to check, separately from the model's answers, whether information survives image downscaling, whether the API roles fit the implementation, and whether results can be reproduced with fixed tool settings.
The real test will be how closely an independent re-run, using the same harness and tool conditions, can reproduce these published figures. Only once latency, throughput, and real-world visual accuracy are demonstrated can anyone judge whether this 384-token design is ready to be extended to production-grade visual decision-making.