IQuest-Q1: Open 320B MoE model built for coding agents on the command line

What it's about

On October 22, 2026, the paper "IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems" was submitted to arXiv (Computation and Language category). It describes a new language model built specifically for agentic coding tasks on the command line (CLI). "Agentic" means the model doesn't just answer questions. It carries out actions on its own over multiple steps, such as running terminal commands, editing files or running tests.

According to the authors, IQuest-Q1 is a sparse mixture-of-experts (MoE) model with 320 billion parameters in total, of which only 15 billion are activated per token. Parameters are, roughly speaking, the knobs the model has tuned during training. In an MoE, the network consists of many specialized sub-networks ("experts"), and a small router picks only a few of them for each token. That gives the model a lot of capacity while using only a fraction of it per token.

The only source is the abstract. It gives no details on architecture, benchmarks, license or download location, and it doesn't mention a Hugging Face repository.

The three-stage training

The abstract describes three phases:

1. Stable and efficient pre-training. The authors combine several techniques: - Stable SplitNorm: a normalization variant meant to keep training stable (the abstract gives no details). - Muon: an optimizer, i.e. the method that determines how parameters are adjusted after each step. - muP (Maximal Update Parametrization): a method for transferring hyperparameters such as the learning rate from small models to large ones.

On top of that comes a two-stage curriculum, i.e. a fixed ordering of the training data. In the second stage, the share of code data is increased. The authors say this preserves a wide stability margin between "safe" and "optimal" learning rates. That matters in practice, because when training large models it's easy to pick a learning rate that makes training unstable.

2. Fine-grained mid-training. Between pre-training and final polishing, the authors broaden the data mix across domains and formats and gradually extend the context, meaning the amount of text the model can process at once.

3. Post-training across multiple scaffolds and versions. It combines: - SFT (supervised fine-tuning): the model learns from examples of good answers. - Domain-specific RL (reinforcement learning): the model is rewarded for succeeding in particular areas, such as solving coding tasks. - MOPD: the abstract gives only the acronym and doesn't spell it out. What exactly it stands for isn't clear from the source.

Training runs across different "scaffolds." A scaffold is the software shell around the model that provides tools and controls the workflow (for example, a coding agent in the terminal). Training with several scaffolds is meant to keep the model from working well in only a single environment. The authors speak of "complementary training signals" that are brought together this way.

Model merging

The authors also apply model merging: experts and checkpoints (saved intermediate states) from different training phases are fused into one model. The paper stresses that this is meant to ease conflicts between the respective parameter updates. The abstract doesn't say how exactly this works.

Results and availability

The abstract only says the model shows capabilities on both coding and general tasks. Concrete benchmark numbers, comparisons with other models and context length are missing. IQuest-Q1 is to be released as an open-source model to support research and development in the community. Whether and where the weights are already available isn't clear from the source.

Our take

The model fits the trend of training large MoE models specifically for agentic coding. With 15 billion active parameters, the compute load per token would be moderate. But the memory needed to hold all 320 billion parameters stays high, so running it on a home PC would be barely feasible without heavy compression. The focus on training stability and on merging multiple training signals is interesting.

A caveat: without benchmarks, license details and weights, quality can't be judged right now. All statements about "capabilities" are vendor claims from the abstract. Wait for the full paper and the actually released weights before drawing conclusions.