IQuest-Q1

Coding agents that run for many steps and call tools along the way are the target of IQuest-Q1, an open-weight model from IQuestLab (IQuest Research), the group that also makes the IQuest-Coder models. The weights reached Hugging Face on 28 September 2026, and a GitHub repository was opened that day. IQuestLab describes it as built for agentic coding, reasoning and multi-step tool use.

Architecture

The design is a mixture of experts holding about 320 billion parameters, with 15 billion in use per token (8 of 256 experts). Its 88 layers alternate three sliding-window attention layers, with a 4,096-token window, and one full-attention layer, which keeps long inputs affordable. Its context runs to 524,288 tokens. Input is text only, in English and Chinese, and added multi-token-prediction layers make decoding faster.

Running it yourself

The model card gives ready configurations for serving with SGLang or vLLM. It names no hosted API, so use means running the weights on a business's own hardware.

A licence condition to note

The licence is MIT with one change. A commercial product or service that builds on the model, or on a derivative of it, has to show the name IQuest-Q1 clearly in its interface. A business planning to embed it should check that requirement against its product design.

Reported results

IQuestLab publishes results on DeepSWE v1.1, Agents' Last Exam, CyberGym and Terminal-Bench 2.1, plus its own IQuest-CLIBench, in a chart on the model card. These are IQuestLab's own measurements. Its model card and code repository are the only sources, and no independent evaluation had been published when the details were last checked on 29 September 2026.

IQuest-Q1 is an open-weight model from IQuestLab, the team behind the IQuest-Coder models. It is aimed at coding agents, reasoning and tasks that chain several tool calls. It was published on Hugging Face on 28 September 2026, and every benchmark result is IQuestLab's own.

What IQuest-Q1 does

• Agentic coding: aimed at coding agents that carry out several steps, including tool calls.
• Long context: a 524,288-token window, with full-attention layers spaced among sliding-window layers to keep long inputs affordable.
• Mixture of experts: about 320 billion parameters, with 15 billion active per token across 8 of 256 experts.
• Faster output: multi-token-prediction layers speed up decoding.
• Self-hosting: the model card gives configurations for SGLang and vLLM, and names no hosted API.

Clients

Teams that want to self-host a text model in English and Chinese for coding and tool use.

Key facts

• Developer: IQuestLab (IQuest Research).
• Published: 28 September 2026 on Hugging Face.
• Languages: English and Chinese text only.
• Licence: modified MIT. Any commercial product built on the model must prominently display IQuest-Q1 in its user interface.