IQuest-Q1

Not yet independently verified. IQuestLab’s own model card and code repository, both created on 28 September 2026, are the only sources; no announcement by the developer and no dated independent report were found. Every benchmark figure is IQuestLab’s own, including an in-house test. We will update this when it can be confirmed, and remove this note.

IQuest-Q1 is an open-weight model from IQuestLab (IQuest Research), the team behind the IQuest-Coder models, published on Hugging Face on 28 September 2026 with a GitHub repository created the same day. It is built for agentic coding, reasoning and multi-step tool use. It is a mixture-of-experts model with about 320 billion parameters, of which 15 billion are active for each token (8 of 256 experts), across 88 layers that alternate three sliding-window attention layers (a 4,096-token window) with one full-attention layer, which keeps long inputs affordable; the context length is 524,288 tokens. It takes text only, in English and Chinese. Multi-token-prediction layers are included for faster decoding, and the model card gives ready configurations for serving it with SGLang or vLLM; no hosted API or price is named. The licence is MIT with one change a business must note: any commercial product or service built on the model, or on anything derived from it, must prominently display “IQuest-Q1” in its user interface. IQuestLab publishes results on DeepSWE v1.1, Agents’ Last Exam, CyberGym, Terminal-Bench 2.1 and its own IQuest-CLIBench in a chart on the model card; those are its own measurements and are not stored as scores, and no independent evaluation had been published when this was read on 29 September 2026.