DeepSeek shows rapid vision gains: agent self-checks screenshots to build 3D scene
cedric_chee · x · 2026-10-17
DeepSeek says vision-related capabilities are advancing rapidly since vision was added to its flagship model, letting anyone describe an idea and build their own games. Developer cedricchee verified a detailed visual-reasoning trace, initially suspecting reward hacking before confirming it was genuine. The screenshot shows DeepSeek-V4.1-Flash building a voxel pagoda in 18m59s—48 steps, 234 tok/s, 5.6M tokens, 99.5% cache hit—reading its own preview screenshots, flagging issues like washed-out contrast and a floating pagoda, then iteratively repositioning trees, waterfalls, and colors.
More from coding & agent
- Pydantic Monty ships TypeScript bindings to sandbox model-written Python in Node and browsers — samuelcolvin · 2026-10-17
- LangChain's deepagents adds Gemini support with 1M tokens and agentic video understanding — hwchase17 · 2026-10-17
- Dev swaps codex remote and Claude remote control for T3 Code mobile — rudrank · 2026-10-17
- Perplexity Flash matches Wald's accuracy at nearly 2x speed in local agent-decision benchmark — WolframRvnwlf · 2026-10-17
- Celesto CI runs cache-hit jobs 23% faster than GitHub Actions on half the CPUs — aniketmaurya · 2026-10-17
- Jev vs tabular foundation models: zero-shot decision routing is not grounded prediction — Arindam_1729 · 2026-10-17