IQuest-Q1

IQuest-Q1 is an open-weight MoE LLM from IQuestLab: 320B total / 15B active parameters, 512K context, built for agentic CLI coding (Claude Code, Codex and similar).

Key facts

• Architecture: 320B-total MoE, ~15B active per token
• Context: 512K tokens
• Benchmark: 83.2 on Terminal-Bench 2.1
• Technical report: released Oct 22–23, 2026 (HF Papers 2610.30949)
• Weights: IQuestLab org on Hugging Face / ModelScope

Why it matters

Small active size plus long context suits terminal agents that read/write files and run commands over many turns — a new contender among open coding models.

How to use it

1. Download weights: huggingface.co/IQuestLab/IQuest-Q1 (or ModelScope)
2. Serve with SGLang (cookbook available) behind an OpenAI-compatible endpoint
3. Plug it into Claude Code / Codex-style CLI agents as the backend model

Project site: iquestlab.github.io.

320B weights require multi-GPU serving; verify scores against the technical report.