IQuest-Q1 released as open 320B MoE model for agentic coding
A new paper presents IQuest-Q1, a sparse Mixture-of-Experts (MoE) agentic foundation model with 320B total parameters, of which 15B are activated. The authors say they release it as an open-source model to support community research and development.
The paper describes a three-stage development process. The first stage is pre-training, which the authors call stable and efficient. It combines Stable SplitNorm, Muon and muP with a two-stage curriculum, and the second stage of that curriculum increases the share of coding data in the mixture. The authors say this setup preserves a wide stability margin between safe and optimal learning rates.
The second stage is what the paper calls fine-grained mid-training. It expands the data mixture across domains and formats and progressively extends the context.
The third stage is multi-scaffold, multi-version post-training. It combines SFT, domain-specialized RL and MOPD across various scaffolds, with the aim of integrating complementary training signals and strengthening agentic coding capabilities.
On top of the three stages, the authors use model merging. It combines different experts and checkpoints from different stages while mitigating conflicts among their parameter updates.
The abstract says the resulting IQuest-Q1 demonstrates capabilities in both coding and general tasks. It states this without figures.
Key facts
- IQuest-Q1 is a sparse Mixture-of-Experts agentic foundation model with 320B total parameters (15B activated).
- Development has three stages: pre-training, mid-training, and multi-scaffold post-training with SFT, domain-specialized RL and MOPD.
- Pre-training uses Stable SplitNorm, Muon and muP with a two-stage curriculum; the second stage raises the coding-data share.
- Model merging combines experts and checkpoints from different stages while mitigating conflicts among their parameter updates.
- The authors release the model as open source; the abstract reports coding and general capabilities without numbers.
Why it matters
Open-weight models aimed at agentic coding are a specific niche. IQuest-Q1 adds a large sparse model to it: 320B total parameters with only 15B active per token, a ratio typical of MoE designs. The paper also documents a full recipe, from optimizer choices in pre-training to merging at the end, which gives other teams a description of how such a model was assembled.
Who it affects
Researchers and developers who work on coding agents and open models are the intended audience; the authors say the release is meant to support community research and development. Teams that train their own large models may find the pre-training choices (Stable SplitNorm, Muon, muP) and the merging approach of interest.
How to use it
The abstract names no license, download link or repository, so there is nothing to point to yet for running the model. The paper is the place to start for the training recipe: the staged data mixture, the multi-scaffold post-training and the merging step.
How solid is it
This account rests on the paper's abstract. The training process is described in some detail, but the abstract gives no benchmark results and no comparison with other models. The claim that the model shows capabilities in coding and general tasks is stated without figures, so it cannot be checked from this text.
Risks and caveats
Without benchmark numbers, the strength of the model is unproven here. The abstract does not name authors or institutions, and it gives no context length, training-token count or compute cost. It also does not define Stable SplitNorm or expand the acronyms SFT, RL and MOPD. Treat the release as promising but unverified until the full paper and independent evaluations are read.