Hands-On: ByteDance's New Doubao 2.1 Pro Turns Floor Plans and Sketches into Working 3D Apps
机器之心 · wechat · 2026-09-18
Jiqizhixin tested ByteDance's newly launched Doubao 2.1 Pro (build 0915, now fully live on Volcano Ark and in Doubao Workspace). Official update pillars: multimodal coding, agentic task delivery, 3D/professional document understanding, and token efficiency. In tests, the model produced an interactive 3D Shanxi travel guide from a single prompt, rebuilt a Little Prince castle in 3D from five-view drawings, coded a playable Flappy Bird from folding-phone demo videos, converted architectural floor plans into interactive 3D models, and built a multi-theme personal dashboard from a hand-drawn sketch. ByteDance says it ran 36 hours of sub-agents on the open-source game Luanti (387K lines), with 83% of fixes merge-ready, and understood 280K lines of legacy Java from just screen recordings and sketches. Image/video inference token usage dropped over 30% versus the previous generation.
Related event: Doubao 2.1 Pro 0915 Update Boosts Agent and Coding Abilities(3 posts)→
More from coding & agent
- Human-in-the-Loop theater: Devs rubber-stamp Claude's 400-line 'dangerous' bash scripts — rickasaurus · 2026-11-01
- One Claude Code command audits your Managed Agents setup against Anthropic's security guide — brada · 2026-11-01
- Claude-Code-Game-Studios: open-source template turns Claude Code into a 49-agent game studio — lxfater · 2026-11-01
- OpenMoat: open-source local policy checks to secure AI coding agents with machine access — __dotk__ · 2026-11-01
- Senior engineer: 'popularity is the weights of an LLM' — half of 600 engineering scars can't be baked into models — RandalSchwartz · 2026-11-01
- Factory's agent ships an MCP connector registry; david hoang wants Paper added — davidhoang · 2026-11-01