IQuest-Q1: 320B MoE model generates games from one prompt and debugs RL training
量子位 · wechat · 2026-09-29
IQuestResearch released IQuest-Q1, a decoder-only Transformer with sparse MoE architecture (320B total, 15B active parameters).
Two hands-on tests from the author:
- One-prompt game generation: a 200-character prompt produced a playable vertical-screen HTML game with NPC design, health bars and random reward mechanics.
- Real engineering debugging: using the ClaudeCode CLI framework, the model analyzed RL training logs and trajectories, locating a bug where the inference server inserted an extra space during text decoding—breaking prefix matching and multi-turn generation—and fixed it.
On benchmarks, it performs well among same-scale models on NL2Repo, CyberGym, Terminal-Bench 2.1, DeepSWE v1.1 and JobBench. The team's earlier ModularRSI self-evolution framework hit #2 on the HuggingFace daily leaderboard, lifting Agent accuracy on Terminal-Bench 2.0 and SWE-Bench Verified from 47.57%/73.40% to 52.43%/76.45%. The model is also being used in its own R&D iteration loop.
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- Fable 5.2's Puzzle Skills Look Suspiciously Good, Reciting Outdated MazeBench Solution — patience_cave · 2026-11-16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