As of Beijing time October 23, 2026, IQuest Research published IQuest-Q1 on its website, with weights on Hugging Face as IQuestLab/IQuest-Q1 and a companion paper on arXiv (2610.30949). The model card lists roughly 320B total parameters, 15B activated per token, 256 experts with 8 activated, and a 524,288-token context window.
The How to use section walks through downloading weights, serving an OpenAI-compatible endpoint with SGLang or vLLM, then pointing Claude Code or Codex at a gateway that supports tool calling. Recommended harness versions on the site are Claude Code `2.1.140` and Codex CLI `0.142`. The `IQuest-Q1[1m]` suffix is described as a client-side setting that does not raise the model’s 512K context limit.
On evaluation, the project reports scores across coding and general-agent benches under its own harness notes. Those figures are author-reported under the published setup—not independent third-party runs. Limitations on the website and model card are explicit: text-only input with no native image, audio, or video; outputs and tool calls need review; real CLI work still needs human oversight; and the project describes itself as early-stage with substantial limits. Hugging Face tags the license as `other`—that is not Apache or MIT without a separate license text. Related reading: Claude Code 2.1.310 Opens Monitor Tool on Bedrock, Vertex and Foundry, Codex for Open Source Offers Maintainers Six Months of ChatGPT Pro, and Kimi Adds K3.1 Pro.
As for how well Q1 holds up on your repos inside Claude Code or Codex, what GPU footprint the eight-way serve actually needs, and whether the reported benches match your tasks—we don’t have that information yet. If you’ve tried it, you can send feedback through the site-owner contact link in the footer. Open weights and a public recipe are not the same as a free hosted agent API.