Chinese team open-sources 320B MoE model IQuest-Q1 that debugged its own training
新智元 · wechat · 2026-09-30
Chinese team IQuestResearch released IQuest-Q1, a 320B open-source sparse MoE model with weights on GitHub and Hugging Face.
- Self-debugged training bug: during RL training the score plateaued with no errors; the model itself dug through logs and code and found the inference server inserted an extra space into outputs, breaking the system's char-by-char history matching and truncating most of the conversation from training. Fixing it lifted average score from 0.704 to 0.769 (0.799 late-stage).
- Training: multi-teacher online policy distillation (MOPD) with four teacher models grading the student's own mistakes, then model merging.
- Benchmarks: strong results on NL2Repo, CyberGym, Terminal-Bench 2.1, DeepSWE v1.1, JobBench.
- Hands-on tests: a 6-intersection traffic-signal task (green-wave coordination; red-light wait cut from 6 min to 22 s), warehouse robot dispatch (87% of 420 Black Friday orders delivered, zero collisions), and elevator scheduling (evening punctuality 26%→61%). Front-end demos include a 2,000+ line single-file animated scroll painting and an ink-style bullet-hell game.
Related event: IQuest Open-Sources IQuest-Q1, a 320B MoE Model for Agentic Coding(11 posts)→
More from Models
- Fable 5.2 Leads MazeBench Token Efficiency; Billion-Token Runs Underway — patience_cave · 2026-11-16
- Fable 5.2's Puzzle Skills Look Suspiciously Good, Reciting Outdated MazeBench Solution — patience_cave · 2026-11-16
- Fable 5.2 Tops MazeBench With 21% After 90 Hours and $6,000 in Credits — patience_cave · 2026-11-16
- Tencent Hy4 flash and InternLM Sage-S2.1-Flash both released, tool-calling tests requested — MajesticPuzzler · 2026-11-16
- User Shares Coral Reef Creation Labeled 'Qwen 4 Max (xhigh)', an Unannounced Model Name — intellectronica · 2026-11-16
- Stealth Model 'silicon-2' Surfaces in Devin, Suspected to Be Anthropic's Next Model — kimmonismus · 2026-11-16