IQuest-Q1: What to Know About the Coding-Agent Model

IQuest-Q1 is an open-weight model from IQuest Research designed for command-line coding agents: systems that inspect repositories, use tools and work through multi-step software tasks. It combines a 320B-parameter sparse mixture-of-experts architecture with 15B active parameters per token and a reported 512K-token context. The design and training target agent workflows, but published benchmark scores alone do not establish how reliably it will complete a particular CLI task. 26815

Release date and licenses

The Hugging Face model page lists the model with the license identifier iquest-q1, and reporting says the weights and model card appeared on September 28, 2026. IQuest’s public release was reported the following day, September 29. The GitHub repository also provides inference code, though the available sources do not establish a separate release date for that code. 456

The model-weight license and the inference-code license should not be conflated. The GitHub repository calls its software license Modified MIT and adds a condition: commercial products or services using the software or derivatives must prominently display “IQuest-Q1” in the user interface. The model card, meanwhile, uses the separate iquest-q1 license label. Review the applicable terms for each component before incorporating them into a product. 420

Architecture and context window

Q1 uses a decoder-only Transformer with a sparse mixture-of-experts design. The model has about 320 billion total parameters, with about 15 billion activated per token; its configuration lists 256 experts, of which eight are activated at a time. The model page also describes a hybrid attention pattern that mixes sliding-window and full-attention layers. 26

Its reported context window is 524,288 tokens, commonly described as 512K. That capacity can accommodate long prompts and repository context, but it does not make the model a lightweight deployment: the total model still requires substantial serving capacity, and practical memory and latency depend on the chosen setup and context length. 1214

Serving and CLI integration

SGLang and vLLM are the serving options cited in the available deployment material. The guidance describes multi-GPU tensor parallelism, with an eight-way example, rather than ordinary single-GPU local use. vLLM also reported day-one support for Q1, including support for its mixed attention pattern. 114

The model is described as integrating with command-line coding tools including Claude Code and Codex CLI. In practice, teams should validate tool-call formatting, memory use and latency with their own agent harness and working context length; a 512K maximum context does not by itself establish how a deployment will perform. 3

Training for agent tasks

IQuest describes Q1 as trained for coding, reasoning, tool use, long-context understanding and multi-step execution. Development accounts also describe multi-harness reinforcement learning and multi-teacher on-policy distillation (MOPD), an approach intended to train the model through its own interactive trajectories rather than only through static example answers. 6924

Those methods explain the model’s agent-oriented design, but they are not a guarantee of dependable behavior in every tool environment. Results still depend on the harness, task and serving configuration.

What the published benchmarks show

A ModelScope post lists scores of 84.5 on CyberGym, 83.2 on Terminal-Bench 2.1, 64.6 on DeepSWE v1.1 and 63.0 on NL2Repo. These benchmarks cover different types of work, including security tasks, terminal use, longer coding tasks and repository generation. The figures are published results, not independent evidence that Q1 will outperform another model in a given CLI setup; comparisons require matching benchmark versions and evaluation conditions. 1521

Q1’s role in its own development

Accounts of IQuest’s training describe researchers asking Q1 to help investigate a problem in the training pipeline. The model analyzed logs and recorded interaction traces to help identify a text-formatting issue in how earlier turns were passed back through the system. Researchers reviewed the diagnosis; the account illustrates supervised development assistance, not a model autonomously improving or releasing itself. 2728

What the evidence means for coding-agent use

Q1’s architecture, long context and training focus make it a candidate for supervised, tool-using coding workflows. Its scale means deployment is not trivial, and benchmark scores should be treated as reported results rather than a promise of performance on a specific repository.

For practical use, run it in a controlled environment, verify changes with tests and keep human review for consequential code or commands. Those safeguards are sensible for any coding agent; the available sources do not establish that Q1 is reliable enough for unattended repository maintenance.