Four months after shipping its V4-Pro and V4-Flash flagship pair, DeepSeek has given its budget-speed line a pair of eyes. On August 21, DeepSeek launched V4-Flash-Vision-Exp, an experimental multimodal model that layers image understanding on top of DeepSeek-V4-Flash's text capabilities. It is positioned not as a new flagship but as infrastructure for agent applications — combining visual understanding with tool use for workflows that need to see while they act.

The Scorecard: Two Wins, One Still Chasing

According to DeepSeek's own comparison data (compiled by OfficeChai), the new model lands close to the previous V4-Flash-0731 on text-based agent evaluations, supporting the company's claim that the vision variant doesn't sacrifice text performance. On Terminal Bench 2.1, V4-Flash-Vision-Exp scores 83.9 versus 82.7 for the older V4-Flash, just 1.1 points behind Opus 4.8's 85.0. But the gap widens on NL2Repo — 57.7 against 69.7 — and DSBench-Hard trails Opus 4.8 by roughly eight points.

The multimodal side is where the release earns its billing. On ApexBench (Pass@1), the vision model scores 36.5, a jump of more than ten points over the 26.2 the text-only V4-Flash manages when forced to ignore image inputs — though still behind Opus 4.8's 39.4. On Agents' Last Exam (27.3 vs 25.7) and ZeroBench Pass@5 (35.0 vs 34.0), V4-Flash-Vision-Exp actually edges past Anthropic's current flagship. On Chartography, 64.3 vs 65.0 is nearly a tie.

One caveat: these numbers come from DeepSeek's own Harness Minimal Mode setup (top_p 0.95, temperature 1.0), not independent verification — treat them with appropriate caution.

Engineering Details Matter More Than Scores

The release also ships developer-facing infrastructure (official Vision guide):

  • Content-based format detection: the model handles JPEG, PNG, GIF and WebP, determining the format from actual file bytes rather than filename or declared MIME type;
  • 384-token cap per image: each image costs at most 384 tokens at V4-Flash's existing rates; an optional detail field downscales images to 512 x 512 to save tokens, with automatic normalization to roughly 800 x 800 before processing;
  • Free Files API: upload an image once and reference it by ID across requests (64 MiB per file), with up to 600 images per request and 8,192-pixel max edge length;
  • Triple-protocol support: OpenAI's Chat Completions and Responses APIs plus Anthropic's Messages endpoint all work, and Harness 0.1.1 released the same day supports the new model out of the box.

An Experimental Model Is a Cheap Move

OfficeChai's analysis cuts to the strategic core: rather than training a new flagship from scratch, DeepSeek added vision to the already-cheap V4-Flash, extending its lineup into multimodal agent territory at minimal cost. The same playbook is visible at MiniMax and Zhipu AI — multimodal capability is sinking into mid-tier models instead of remaining an exclusive selling point for giant flagships.

The takeaway for developers is direct: when a cheap model that can see approaches flagship-level multimodal benchmarks, model selection shifts from "who scores highest" to "visual agent capability per unit cost." The experimental label means it could be adjusted or graduate at any time — sensible for low-risk workflow trials, not yet for betting the production critical path.

The next round of competition among Chinese AI labs may not be about who has the biggest model, but who can teach a cheap, fast one to see the world first. DeepSeek has played its card.