Chen Danian's StartLux: 27B local model ranks 2nd in CAICT MCP test, near DeepSeek-V4-Pro
量子位 · wechat · 2026-09-03
Chen Danian, co-founder of Shanda, is back with StartLux, a local-model startup. Its first model, StartLux-V1.0-27B-Preview, scored 39.25% in CAICT's MCP agent benchmark — 2nd overall, just 1.3 points behind the 1.6T-parameter DeepSeek-V4-Pro and 5.34 points above same-size Qwen3.6.
Built on Qwen3.6-27B and post-trained with an 'AutoResearch' AI-trains-AI method focused on tool calling and multi-step execution, it runs locally on consumer PCs. In hands-on tests it beat Claude Sonnet 4.6 on financial calculation accuracy and browser flight search (95s to a $299 fare vs 200+s and $556). Chen predicts local models will match Claude and take 80% of the cloud market within three years; a one-click local solution ships this year.
More from Models
- OpenAI's Astra Said to Autonomously Find and Exploit Security Vulnerabilities — VraserX · 2026-09-03
- User Slams GPT-5.6 Max Reasoning for Getting 'Stupid' Overnight, Speculates GPT 6 Astra Is Near — CtrlAltDwayne · 2026-09-03
- Opus 5 all-day usage report: no parallel subagents means limits are a non-issue — 4310sy · 2026-09-03
- Mathematician writes human-readable digest of Claude's 2/3 zeta zeros proof — Thom_Wolf · 2026-09-03
- Recurrent Depth Debate: NVIDIA's Deja Vu Nears 1B-Param Performance with 10M Params — ZGojcic · 2026-09-03
- Wes Roth Builds Four Full AI Games on Claude Fable 5.1's Low-Effort Setting — Wes Roth · 2026-09-03