Deployment considerations for research
A.X K1 is a high-capacity Mixture-of-Experts language model designed to balance large-scale reasoning with practical inference cost. With 519B total parameters and 33B active parameters per token, 131K token context, and 61 layers (60 MoE + 1 dense), it enables deep reasoning while maintaining feasible latency for interactive use. The Post-MLP RMSNorm further stabilizes training, and the Think/Non-Think modes give users control over the depth of reasoning versus speed.
Evaluations for A.X K1 are scheduled for January 4, 2026, and the model is intended for research-grade deployments with careful resource planning. Running locally with SGLang and vLLM is supported, but practitioners should note possible limitations like potential hallucinations and domain gaps. Cite the technical report if you publish results.