IQuest open-sources IQuest-Q1, a 320B MoE model for agentic coding

IQuest Research has open-sourced IQuest-Q1, a sparse mixture-of-experts model with about 320 billion total parameters and roughly 15 billion active per token, aimed at agentic coding and multi-step tool use. Weights and a model card went up on Hugging Face on September 28, with inference code on GitHub under a custom license.

• 320B total parameters, about 15B active per token
• 88 layers, 256 experts with 8 active, 524,288-token context
• Benchmarks: 64.6 DeepSWE v1.1, 83.2 Terminal-Bench 2.1, 84.5 CyberGym
• Recommended serving via SGLang or vLLM on eight GPUs