IQuestLab Ships IQuest-Q1, a 320B Sparse Coding Agent You Can Self-Host
A new open-weight 320B Mixture-of-Experts model from IQuest Lab targets agentic coding in the terminal, activating 15B parameters per token with a 524K context window.
Takeaways
- IQuest Lab released IQuest-Q1, an open-weight 320B MoE model with 15B activated parameters per token.
- Scores 83.2 on Terminal-Bench 2.1 and 57.5 on SWE-Bench Pro for agentic coding.
- Uses 256 experts with 8 routed per token, hybrid sliding-window attention, and 524K context window.
- Ships with native Claude Code and Codex CLI integration, deployable via SGLang or vLLM.
- Trained via Stable SplitNorm, Muon, muP pretraining, then SFT, domain RL, MOPD, and model merging.
- Text-only, no multimodal input, released under a custom IQuest-Q1 license.
IQuestLab has released IQuest-Q1 weights, a 320-billion-parameter Mixture-of-Experts model built for terminal-based coding agents. The model activates about 15 billion parameters per token, supports a 524,288-token context window, and includes documented integration paths for Claude Code and Codex CLI.
The release gives engineering teams another option for running coding agents on private infrastructure. IQuestLab reports competitive results on terminal and repository-level software engineering benchmarks, although those results still require independent replication and careful comparison under matched harnesses.
Benchmarks put the claim in focus
| Benchmark | Reported score | What it measures |
|---|---|---|
| Terminal-Bench 2.1 | 83.2% | Agent performance on tasks completed through a terminal |
| SWE-Bench Pro | 57.5% | Resolution of software issues in real repositories |
Agent benchmarks measure the complete system around a model, including prompts, tool schemas, retry policies, sandbox behavior, and time limits. Comparisons are most useful when every model runs the same benchmark version through the same harness. IQuestLab’s figures show that the downloadable checkpoint can handle demanding agent workflows; they do not remove the need for local evaluation on representative repositories.
Inside the sparse design
IQuest-Q1 combines a large expert pool with selective activation, allowing each token to use a small fraction of the model’s total capacity. Its main architectural specifications are:
| Transformer layers | 88 |
|---|---|
| Hidden dimension | 3,072 |
| Attention heads | 48 query heads and 8 key/value heads |
| Attention pattern | Three sliding-window layers for each full-attention layer |
| Sliding window | 4,096 tokens |
| Expert routing | 8 selected from 256 experts per token |
| Active parameters | Approximately 15 billion per token |
| Maximum context | 524,288 tokens, commonly rounded to 512K |
The first transformer layer uses a dense multilayer perceptron instead of sparse expert routing. IQuestLab says this choice reduces expert-load imbalance near the start of the network. Multi-Token Prediction layers can also support speculative decoding in compatible inference servers by proposing several tokens during each generation step.