IQuest-Q1
320B sparse MoE (15B active) agentic-CLI model with 512K context and 83.2 on Terminal-Bench 2.1.
This entry is marked for review. Check the original sources before relying on its details.
Overview
IQuest-Q1 is a sparse mixture-of-experts transformer with 320B total and 15B active parameters, 88 layers and 256 experts (8 active per token), described in an arXiv technical report dated October 23, 2026 as an open-weight agentic model for command-line and software-engineering work. It interleaves grouped-query attention with sliding-window attention in a 3:1 ratio and was mid-trained progressively to 32K, 128K and 512K context. The paper 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; these are the authors' own figures. The paper page names no license and links no weights, so availability terms are unconfirmed.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 320B total / 15B active parameters, 256 experts with 8 active per token
- 88 layers with 3:1 sliding-window to full-attention interleaving
- Progressive context extension through 32K, 128K and 512K tokens
- Reported 83.2 on Terminal-Bench 2.1 and 57.5 on SWE-Bench Pro
- Reported 85.5 on BrowseComp and 84.5 on CyberGym
- Trained with the Muon optimizer and muP parameterization
Use cases
- Terminal and repository-level agentic coding
- Long-horizon tool-calling workflows
- Research into sparse-MoE agent training recipes
Pricing
unknown: no license or hosted pricing stated on the technical report page.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 23 Oct 2026
- Record updated
- 26 Oct 2026
- Record last checked
- 26 Oct 2026