IQuest-Q1 becomes largest Muon-trained model at 320B parameters

Beijing-based IQuest Research published a technical report on IQuest-Q1, a 320-billion-parameter sparse MoE model it calls the largest known deployment of the Muon optimizer. The company claims 1.93x compute efficiency over a comparable AdamW-trained MoE baseline, though no independent confirmation has been published.

• 320B total parameters, ~15B activated per token, 8 of 256 experts
• Claimed 1.93x compute efficiency vs AdamW MoE baseline at equal loss
• 512K-token context, deployable on eight datacenter GPUs
• No independent benchmark results listed yet on gradually.ai