IQuest-Q1 is a 320B open-weight coding model; its Terminal-Bench claim needs a check

IQuest, a lab with a 177-author team, posted a technical report on October 22, 2026 for IQuest-Q1, an open-weight coding model with 320 billion total parameters of which 15 billion are active per token. Its report claims 83.2 on Terminal-Bench 2.1 and 57.5 on SWE-Bench Pro. Those are the lab's own numbers, and a public leaderboard lists other entries well above 83.2, so "tops Terminal-Bench" needs a qualifier.

What the paper says

The arXiv paper (2610.30949) describes IQuest-Q1 as "a sparse Mixture-of-Experts agentic foundation model." The architecture has 88 transformer layers and 256 routed experts, with 8 activated per token, a sliding window of 4,096 tokens and a 160,000-token vocabulary. Training used 29 trillion tokens across pre-training and mid-training, with context extended in steps from 32K to 128K to 512K tokens. Post-training combined supervised fine-tuning, domain-specific reinforcement learning and multi-teacher on-policy distillation across several agent scaffolds, then model merging.

The authors say the model is open-weight and that they release weights, deployment configurations and agent-integration instructions. The Hugging Face paper page we read did not state a license or include a download link, so builders should check the official release before assuming terms. For a primer on how a model like this gets its low per-token cost, see our explainer on mixture of experts and active parameters.

The reported scores

From the paper's selected benchmarks:

• SWE-Bench Pro: 57.5
• DeepSWE v1.1: 64.6
• CyberGym: 84.5
• Terminal-Bench 2.1: 83.2
• GPQA Diamond: 93.3
• Humanity's Last Exam: 39.2 without tools, 55.7 with tools
• BrowseComp: 85.5
• WideSearch: 81.8

The paper also reports a cybersecurity deployment in which the model produced 30 vulnerability reports over three days, all independently validated, and a scaling-efficiency result of about 1.93 times a Qwen3-MoE-style baseline, meaning roughly 52% of the compute for the same training loss. The efficiency figure compares against the authors' own baseline.

Does it really top Terminal-Bench?

Some roundups described IQuest-Q1 as topping Terminal-Bench 2.1 at 83.2. We could not support that. BenchLM's Terminal-Bench 2.1 table, updated October 23, 2026, lists scores from 87.9% to 92.8% for its top ten entries, all above 83.2. BenchLM warns that "harness and effort settings vary by row" and that results are mostly provider-reported, and it keeps Terminal-Bench 2.1 apart from version 2.0 because they are not comparable. IQuest-Q1 does not appear in that top ten.

So there are two honest statements. First, 83.2 is a strong number for a model with 15 billion active parameters, if it holds up. Second, a self-reported figure with an unspecified harness is not a ranking. Until an independent evaluator such as Vals AI or Epoch publishes a run with a stated scaffold, treat the leaderboard position as unestablished. Terminal-Bench scores swing by several points depending on the agent wrapper, so comparing across tables is how bad headlines get written.

Why a 15B-active model matters

Only 15 billion of the 320 billion parameters work on each token, so serving cost looks closer to a mid-size dense model than to a giant one, though all 320 billion must still sit in memory. That is the practical appeal of open-weight sparse models: a team with a few high-memory GPUs can host a model that is competitive on agentic coding, subject to the benchmark caveats.

The paper's other notable claim is how it was trained. The authors credit a data curriculum that raises the share of coding data in the second pre-training stage, and a post-training recipe that distills from several teacher models. Neither is verified by outside reproduction. With 177 authors listed, it is a large industrial effort, not a garage release, and the arXiv page lists a corresponding author but no company details we could check.

How to read a self-reported benchmark

A lab's own table is a claim about a specific setup. For agentic coding, the setup includes the scaffold that wraps the model, the number of attempts allowed, the time limit per task and the tools exposed. Change any of them and a Terminal-Bench score can move by several points. That is why independent leaderboards record the harness, and why BenchLM separates versions of the same benchmark. The IQuest paper lists post-training "across different agent scaffolds," which is useful for portability but makes a single headline number harder to interpret. A fair comparison runs every model through one scaffold, with the same budget, on the same day.

The security result deserves a footnote

Thirty validated vulnerability reports in three days is the kind of figure that drives policy debates about open-weight cyber capability. We covered the gap between open and closed models on that question in our tracker of open-weight cyber capability. The paper's number is self-reported, does not say which software was tested, and we have no third-party confirmation. It is a reason for an independent check, not a conclusion.

What to do with it

If you run coding agents, the model is worth a test on your own repositories once weights are available, using the same harness you use for your current model. If you evaluate models for a procurement or policy team, ask three things: which harness produced 83.2, whether an independent run exists, and what license the weights carry. Ignore the word "tops" until the first is answered.

Our take

IQuest-Q1 is a credible entry in the open-weight coding race and a good test of how quickly outside evaluators can verify a lab's own claims. We would not plan around it yet. Watch for the weights release, the license text and an independent Terminal-Bench run. If those land and the number survives, the 15B-active design becomes the cost story of the quarter.