IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems

What it claims

IQuest-Q1 is a sparse Mixture-of-Experts model with 320B total and 15B activated parameters, aimed at agentic coding in CLI systems. The authors say it is capable in both coding and general tasks, and they describe it as an open-source release. The page gives no benchmark figures to support the capability claim.

How it works

The abstract gives a reason for the pre-training recipe that the summary leaves out. Stable SplitNorm, Muon and muP are combined to keep a wide stability margin between the safe learning rate and the optimal one. The second curriculum stage then raises the share of coding data.

Mid-training widens the data mixture in two ways, across domains and across formats, while extending context length step by step. The page does not state the final context length.

Post-training is "multi-scaffold with multi-version." SFT, domain-specialized RL and MOPD run across several agent scaffolds, so that their complementary training signals combine. The abstract does not expand MOPD or name the scaffolds.

Model merging is the last step. It combines different experts with checkpoints taken from different training stages, and it is designed to reduce conflicts between their parameter updates.

flowchart LR
  A[SFT] --> D[Post-trained variants]
  B[Domain RL] --> D
  C[MOPD] --> D
  S[Multiple scaffolds] --> D
  D --> M[Merge: experts + stage checkpoints]
  P[Pre/mid-training checkpoints] --> M
  M --> Q[IQuest-Q1]

Limits

The page shows only the abstract. It reports no benchmarks, baselines, training compute, context length, licence or download location. The release is called open-source, but the Hugging Face page links no models, datasets, Spaces or collections to the paper as of October 2026. Anyone planning to adopt the model has to find the weights and their licence somewhere else.