IQuest-Q1 opens 320B/15B MoE, 83.2 on Terminal-Bench 2.1
TL;DR
- IQuest-Q1 activates only 15B of 320B total parameters per token, targeting frontier-class output at near-small-model inference cost.
- Native Claude Code and Codex CLI integration hooks let teams substitute IQuest-Q1 without rewriting their agentic toolchain.
- SGLang shipped day-zero inference support, signaling immediate infrastructure-layer alignment on release day.
IQuest released IQuest-Q1 as an open-weight sparse mixture-of-experts with 320B total and 15B activated parameters, 88 transformer layers, 256 experts (8 routed) and a 512K context window. The paper, Advancing Foundation Capabilities for Agentic CLI Systems, reports 83.2 on Terminal-Bench 2.1, 57.5 on SWE-Bench Pro, 64.6 on DeepSWE v1.1, 93.3 on GPQA Diamond and 85.5 on BrowseComp.
The headline figure sits inside a wider results table that is less uniformly flattering. On the newer Terminal-Bench 3.0 the authors list 17.9, and 22.7 on Terminal-Bench 4.0, well below the 83.2 score the release leads with. CyberGym comes in at 84.5.
The training recipe is described as a three-stage process: "stable and efficient pre-training," "fine-grained mid-training" that progressively extends context to 512K, and "multi-scaffold with multi-version post-training" combining SFT, domain-specialized RL and model merging. On compute, the authors claim "1.93× higher scaling efficiency than the baseline, requiring 1/1.93 ≈ 52% of the training compute to reach the same training loss." The paper does not name the baseline model that figure is measured against.
The release is paired with a security deployment. "During a three-day deployment, human cyber security experts, assisted by IQuest-Q1, produced 30 vulnerability reports," the authors write, adding that each report was independently validated by other experts and findings affecting open-source projects were responsibly disclosed to their maintainers. The paper does not list the CVEs or the projects involved.
IQuest-Q1 lands in a stretch of open-weights coding-model drops our tracker is logging almost daily, including Recursive's Flint 40B/2.5B MoE earlier this week.
What others are reporting
-
arXiv (IQuest Research) Read →
First-party technical paper from 177+ authors; defines the three-stage training methodology and releases weights under a custom license for community research.
-
MindStudio Read →
Examines architectural trade-offs and provides explicit Claude Code and Codex CLI integration steps while candidly noting early-stage real-world limitations.
Only about 15 billion are activated per token, thanks to a sparse MoE routing scheme with 256 total experts and 8 activated per forward pass.
-
CCTest Read →
Frames the model as a systems-level training problem, breaking down multi-stage curriculum and domain RL rather than leading with benchmark numbers.
The recipe connects data composition, context growth, RL environments, scaffold design, and checkpoint integration rather than stopping at pre-training plus a tool interface.
-
Pandaily Read →
Straight news coverage emphasizing the open-sourcing decision and the 15B active parameter architecture for a broader tech audience.
Originally reported by huggingface.co
Read the original article →Original headline: IQuest Open-Weights 320B/15B MoE Agentic CLI Model, Hits 83.2% on Terminal-Bench 2.1 and 57.5 on SWE-Bench Pro