DeepSeek-v4-flash-vision-exp
DeepSeek has released an experimental vision-enabled version of its V4 Flash model, prompting interest from developers who rely on it for coding, UI work, and agentic workflows that need to “see” screenshots, documents, and designs. Commenters highlight strong cost-performance and improved autonomy on tasks like frontend QA, OCR, and spatial reasoning, but also note key limitations such as the 800×800 image resize cap, occasional basic vision failures (e.g., reading clocks or landmarks), and concerns over recent pricing changes. Many expect harness-level workarounds (cropping, zoom tools, external vision models) and future model iterations to address fidelity and reliability gaps.
Vision capability and behavior
- DeepSeek V4 Flash previously “acted” like it had vision, trying to process screenshots via text-only tricks; the new vision-exp model is seen as fixing that gap.
- Some note it is still experimental and not a “flagship” vision system, more a fast/cheap coding‑oriented model that now accepts images.
Image resolution limits (800×800) and workarounds
- Images are resized to ~800×800 equivalent, capped at ~384–400 tokens per image; cost estimates suggest ~2,500 images per dollar.
- Many consider 800×800 too low for dense content (A4 pages, schematics, tiny text, detailed diagrams).
- Others argue it’s acceptable for screenshots or small crops and that tooling can compensate:
- Crop/zoom tools or tiling into grids, then stitching results.
- Using separate layout/OCR models (e.g., PP‑DocLayout, Unlimited‑OCR) to split pages.
- There’s debate on whether such harness workarounds preserve global relationships and counting accuracy.
Use cases for vision
- Common “nice” uses:
- Frontend/UI dev loops (screenshots from Playwright/headless browsers, Figma inspection, verifying layouts).
- OCR / scanned docs and CV photos; extracting text while optionally “fixing” or preserving errors.
- Diagrams, charts, design sketches, wireframes, game UIs, learning by sketching and having AI correct.
- Social media alt text, research on satellite imagery, robot perception, 3D printing monitoring, context compaction via tiny-font PNGs.
Quality, benchmarks, and failures
- Benchmarks: reported DeepSWE score slightly above the prior text‑only Flash, with notably better cost/performance vs some competing models.
- Some users praise DeepSeek Flash vs similarly priced models for persistence, validation, and tool use; others report incoherent outputs and poor tool calling (possibly quantization/harness issues).
- Vision accuracy is mixed:
- Fails or inconsistently passes “clock reading” tests; gets some analog times wrong that competing models also misread.
- Performs worse than some rivals on landmark identification (e.g., confusing cathedrals/bridges).
- Several emphasize “jagged intelligence”: strong at coding and many visual tasks, but unreliable on certain edge cases.
Openness, ecosystem, and alternatives
- Expectation (but not confirmed here) that weights may be released, consistent with prior DeepSeek practice.
- Suggestions to attach external vision models via MCP or plugins.
- Some users are switching to other models (e.g., Qwen) due to DeepSeek pricing changes, despite enthusiasm for its capabilities.