IQuest-Q1 open-weights 320B MoE coding agent: 15B active / 524K context; DeepSWE 64.6, Terminal-Bench 2.1 83.2

Background

Open-weight stacks still need models that navigate repos, drive terminals, and finish long tool chains—not just chat leaderboards. IQuest-Q1 from IQuest Research targets agentic coding. Around Oct 8, public materials (PR Newswire, GitHub/HF) highlighted demos and deploy paths with downloadable weights.

Architecture

Sparse MoE decoder-only: ~320B total / ~15B active per token; 88 layers; 256 experts / 8 active; hybrid 3×SWA + 1×FA (window 4096); 2 MTP layers in training, recursive MTP×8 at inference; context 524,288. Post-training mixes SFT+RL with Multi-Teacher On-Policy Distillation (MOPD) on the student’s own rollouts (GitHub README).

Benchmarks

Figures from the official HF performance chart (self-reported / same pipeline). DeepSWE uses mini-SWE-agent; other agentic tasks mostly Claude Code; Agents' Last Exam uses Claude Code 2.1.258 with multimodal inputs replaced by placeholders (text-only checkpoint). CyberGym 6h / Terminal-Bench 2.1 8h caps.

IQuest-Q1: DeepSWE 64.6, NL2Repo 63.0, CyberGym 84.5, Terminal-Bench 2.1 83.2, JobBench 55.7, Agents' Last Exam 29.6, HLE (no tools) 39.2, IQuest-CLIBench 53.7. Vendor notes early-stage reliability limits.

Developer path

hf download IQuestLab/IQuest-Q1, serve via official SGLang/vLLM CUDA 13 images at tp=8 with iquest_q1 parsers (optional MTP). Wire Claude Code 2.1.140 or Codex 0.142 through a compatible gateway. Text-only; custom tool format requires matching parsers.