arxiv.org web signal

Tencent Publishes Hunyuan-A13B Technical Report — 80B MoE, 13B Active, Trained on 20T Tokens With Dual-Mode Reasoning

Open Source China AI ai-business

Summary

Tencent posted the Hunyuan-A13B technical report on arXiv on Sept 23, detailing an 80B-total / 13B-active MoE trained on 20T tokens with enhanced STEM data, high-quality SFT and large-scale RL. The report describes a dual-mode 'fast/slow' reasoning framework that varies compute per query and claims performance approaching much larger models on math, code, general language and agent tasks. Weights ship under Creative Commons Attribution 4.0.