Chinese AI company DeepSeek released DeepSeek-V4-Flash-Vision-Exp on August 21, 2026. The experimental multimodal model is now available on the DeepSeek API platform.

The model extends the text-only DeepSeek-V4-Flash with image understanding. It preserves the base model's performance on reasoning, world knowledge, and agent tasks. On DeepSeek's internal multimodal agent benchmarks, V4-Flash-Vision-Exp scores close to Anthropic's Opus 4.8. This represents a significant leap over the base V4-Flash on vision-dependent tasks. The company also shipped DeepSeek Harness 0.1.1 with out-of-the-box support for the new model.

What's new

  • Model ID: deepseek-v4-flash-vision-exp (experimental)
  • Capabilities: Image description, screenshot text extraction, chart and diagram analysis
  • Supported formats: JPEG, PNG, GIF, WebP (detected from file content, not filename or MIME type)
  • Input methods: Base64 inline, external HTTPS URLs (up to 32 MiB, 8192-character limit), Files API reference (up to 64 MiB per image)
  • API compatibility: OpenAI Chat Completions, OpenAI Responses API, Anthropic Messages endpoint
  • Detail control: Optional detail parameter (low downscales to 512×512; high/original/auto keep full resolution)
  • Token billing: Images resized to ~800×800 equivalent; hard cap of 384 tokens per image at V4-Flash pricing
  • Request limits: 48 MiB body, 600 images max, 64 MiB total without file_id (200 MiB with), 8192 px per side (drops to 4096 px with 15+ images)
  • Restrictions: Images in user messages only; vision models only; reserved placeholder token rejected

Why it matters

The release closes a gap in DeepSeek's toolchain. Developers can now send screenshots and visual context directly to agents via commands like /goal and /plan in Harness. This enables "work by looking at images" workflows without custom glue code. The Files API, now free, lets teams upload an image once and reuse it across requests. That helps with multi-turn debugging or document analysis where re-uploading would waste bandwidth and latency.

Benchmark data from DeepSeek's change log shows V4-Flash-Vision-Exp reaching 63.6 on DSBench-Hard, 64.3 on Chartography, and 35.0 on ZeroBench (Pass@5). The model matches V4-Flash on pure-text agent tasks like Terminal Bench 2.1 (83.9) and NL2Repo (57.7). DeepSeek notes that on Code Agent text tasks, V4-Flash ignores multimodal elements. The vision variant's gains are specific to visual understanding benchmarks.

Our take

DeepSeek is iterating its agent stack in public. Harness 0.1.1-rc.1 arrived the day before with multimodal core support, and the vision model follows immediately. The experimental label and benchmark proximity to Opus 4.8 suggest a capability probe rather than a polished product. Pricing parity with V4-Flash lowers the barrier to test it, but the 384-token image cap and user-message-only restriction will shape what workflows are practical today.

Sources