IQuest Research Open-Sources IQuest-Q1, a 320B MoE Model for Agentic Coding With 15B Active Parameters

IQuest Research has announced the launch of IQuest-Q1, a sparse mixture-of-experts language model designed for coding and AI applications. The model features approximately 320 billion parameters, with around 15 billion activated per token, and the weights and model card were published on Hugging Face on September 28. IQuest-Q1 includes 88 transformer layers and 256 experts, with eight active experts per token and a hybrid attention pattern. It is recommended to run it with SGLang or vLLM across eight GPUs, positioning it as a foundational model for command-line agent systems.