IQuest-Q1 posts a 320B MoE coding agent with a modified MIT license

What did IQuest announce this week?

IQuest Research’s PR Newswire release, datelined Beijing and published 8 October 2026 ET (APAC mirrors list 9 October), says IQuest-Q1 has drawn early praise for coding, software engineering, interactive application generation, and long-horizon agentic workloads. It links GitHub, Hugging Face, and the IQuest blog as the public materials.

The release highlights demos—one-shot interactive apps, debugging an RL training run, and multi-step tool use across office artifacts. Those narratives are IQuest’s; we did not reproduce the demos.

IQuest also posted SAIL earlier as a scientific-agent line; that is a different artifact. See IQuest’s SAIL scientific agent for the prior write-up.

The PR Newswire object is a visibility push pointing to materials that already existed on Hugging Face and GitHub; the dated news for this article remains the 8 October ET release, not a claim that weights first appeared that calendar day. Demo narratives in the release—one-shot interactive apps, debugging an RL run, multi-step office-tool workflows—are IQuest’s storytelling, useful as intended-use hints and useless as independent capability proof.

What is IQuest-Q1 on paper?

The Hugging Face card lists approximately 320B total parameters, 15B activated per token, 88 layers, 256 experts with 8 active, hybrid sliding-window plus full attention, and a 524,288-token context length. Post-training is described as SFT and RL plus Multi-Teacher On-Policy Distillation (MOPD) on the student’s own rollouts.

Sparse MoE routing is the efficiency story. For the architecture pattern without IQuest’s numbers, see our mixture-of-experts explainer.

IQuest’s performance section reports public or same-pipeline scores across NL2Repo, CyberGym, Terminal-Bench 2.1, DeepSWE v1.1, and others, and explicitly recommends temperature 1.0, top-p 0.95, top-k 20 with Claude Code 2.1.140 or Codex 0.142 as harnesses for reproducibility. DeepSWE v1.1 at 64.6 and Terminal-Bench 2.1 at 83.2 are among the figures circulating from those materials—still IQuest’s evaluation story.

Benchmark notes on the card matter as much as the headline scores. IQuest says it uses mini-SWE-agent for DeepSWE v1.1 and Claude Code for other agent tasks, with multi-hour caps such as six hours for CyberGym and eight hours for Terminal-Bench 2.1. Multimodal inputs in agent exams are replaced with placeholders because the checkpoint is text-only. Those protocol choices mean a raw DeepSWE or Terminal-Bench number from another lab’s harness is not automatically comparable.

How do you run it, and what license applies?

IQuest recommends SGLang or vLLM with CUDA 13 prebuilt images and tensor parallel size 8, plus custom `iquest_q1` tool-call and reasoning parsers. Claude Code and Codex are wired through a gateway that speaks Anthropic Messages or OpenAI Responses; the card’s sample Codex config sets approval_policy to never and sandbox_mode to danger-full-access—copy that only if you intentionally accept those risks.

The LICENSE file is titled Modified MIT License. Besides standard MIT terms, it requires that if the Software or derivatives are used for commercial products or services, you shall prominently display “IQuest-Q1” on the user interface. Hugging Face tags the card `license:other` with license_name `iquest-q1`.

That is not the same as an Apache-2.0 or stock MIT grant. For why license text matters when people say “open,” see open weights vs open source.

Serving notes assume serious multi-GPU setups (tensor parallel size 8 in the examples) and custom `iquest_q1` parsers for tool calls and reasoning channels. If you wire Claude Code or Codex through IQuest’s gateway samples, re-read the sandbox and approval flags before copying them into a machine that can touch production systems. The modified MIT mark clause is a product/legal constraint: commercial UIs must prominently show “IQuest-Q1,” which is why Hugging Face tags `license:other` / `iquest-q1` rather than plain MIT.

What limits does IQuest state?

The card’s Limitations section says the checkpoint is text-only, generated code can be incorrect, tool calls need IQuest-specific parsers, real-world CLI tasks may loop or miss constraints, and the model remains early-stage. Multimodal benchmark inputs are replaced with placeholders during tokenization in IQuest’s agent evaluations.

Agentic coding benches are sensitive to harness and time limits. For why SWE-style scores need protocol context, see SWE-bench explained.

Self-hosting a 320B MoE at tp=8 is an infrastructure project. Teams comparing coding agents may also look at harness-centric releases such as Together’s Link coding agents.

IQuest’s own early-stage warnings—incorrect code, looping CLI agents, missing multimodal support—should be the default prior for procurement decks that only paste the 64.6 / 83.2 rows. Treat self-reported agentic coding benches as a starting hypothesis for a bake-off on your repositories, with your latency, cost, and safety envelope measured under your harness.

What did we not test?

We did not download the multi-hundred-gigabyte weight shards, start the SGLang image, or run DeepSWE or Terminal-Bench. This article reports the PR Newswire text, the Hugging Face model card and LICENSE, and the GitHub repository we opened on 10 October 2026.