Hunyuan-A13B runs 13B of its 80B parameters per request

single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC

What this means for you

No decision yet: the report gives architecture and training numbers but no per-benchmark scores, so it cannot be compared against whatever you already serve. The authors claim throughput suits latency-sensitive work; with no figure attached, that is not a reason to migrate. Wait for scores and a harness.

Hunyuan-A13B is an open-weights language model built as a mixture of experts — only part of the network runs on each request, so serving is cheaper than the total size suggests. It holds 80 billion parameters and activates 13 billion per token, pretrained on a 20-trillion-token corpus with extra STEM material, then fine-tuned with reinforcement learning on top. A dual-mode setup writes out more intermediate steps for hard questions and fewer for easy ones. The authors list maths, science, coding, language understanding and agent tasks as evaluated, but publish no scores, dates or harness, and no independent evaluation accompanies the report.

Key facts

  • ·The model holds 80 billion parameters in total and activates 13 billion per token, according to its authors. source
  • ·It was pretrained on a 20-trillion-token corpus with extra curation of STEM data. source
  • ·It uses a dual-mode scheme: shallower step-by-step answering for routine queries, deeper for multi-step problems, in the authors' description. source
  • ·The report lists evaluations in mathematics, science, programming, general language understanding and agent tasks, and publishes no scores, evaluation dates or harness. source
  • ·The authors state the model is released to support open research and deployment; the abstract names no licence or repository. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire