arXiv 2610.30949

IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems

AIOctober 22, 202653 upvotes

Authors: A. Jian Yang, An Huang, Bo Dan, Chen Mu, Chengran Yang, Chuan Shi, Feng Chang, Hongxin Li +169

Overview

The authors present IQuest-Q1, a sparse Mixture-of-Experts agentic foundation model with 320B total parameters (15B activated). It is built through three stages: stable pre-training with a two-stage coding-weighted curriculum, fine-grained mid-training with progressive context extension, and multi-scaffold, multi-version post-training that combines SFT, domain-specialized RL, and MOPD, followed by model merging.

Introduction

Agentic coding systems need foundation models that are stable to train and strong at coding while remaining capable on general tasks. The abstract targets this gap by focusing on training stability and agentic coding ability at scale.

Results

The resulting model demonstrates capabilities in both coding and general tasks, and the authors release it as open source. The abstract does not report specific benchmark numbers.

Conclusion

Because the model and its training recipe are open, engineers and researchers can build on a large agentic coding model without starting from scratch.

Same story, different eyes

The EngineerAI
A wide safety margin on learning rates may be the quiet hero here
Reading this as an engineer, what stands out is the idea of keeping a wide stability margin between the safe and the optimal learning rate. When you train a 320B-parameter MoE, a single unstable run can cost a huge amount of time and compute, so that kind of headroom matters a lot in practice. Merging experts and checkpoints from different stages without their parameter updates fighting each other also sounds genuinely tricky to get right. How much each stage contributes to the final results is not yet known.
The SkepticAI
A wide gap between safe and best learning rates
What caught my eye is the claim of a wide stability margin between safe and optimal learning rates in pre-training. If that holds up, it would explain a lot about how a 320B MoE trains smoothly, so it would be reassuring to see the learning-rate sweeps behind it, and an ablation showing what Stable SplitNorm, Muon, and muP each contribute. Since the open release makes this checkable, I'd also love to see how the merged model scores on coding benchmarks across different scaffolds. How much merging itself adds over the best single expert is not yet known.
The FounderAI
An open 320B agentic coding base model opens the door to scaffold-specific startups
Seen through a founder's lens, the interesting part is that an open-source model was trained across multiple scaffolds and merged from several experts, so it's built to work inside CLI agents rather than just chat. That closes some doors for anyone selling a thin wrapper around a closed coding model, but it opens others. You could start this week by fine-tuning or serving it for one narrow workflow, like repo migrations or test repair in a single language, and measure it against your current closed-model setup. With only 15B parameters active, serving costs may be friendlier than the 320B headline suggests, which is why now is a good time to try. How it holds up on real-world long-running tasks is not yet known.