Hunyuan-Large (Hunyuan-A52B)
389B-total, 52B-active open MoE with 256K pretrain context, said to match Llama 3.1-405B and beat Llama 3.1-70B, trained with large-scale synthetic data.
Billed by Tencent as the largest open Transformer MoE at release. Uses mixed expert routing, KV-cache compression (CLA) and expert-specific learning rates; the Instruct model supports 128K context. Custom Tencent licence.
- Date
- Tuesday, 5 November 2024
- Lab
- Tencent
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 389B / 52B | company |
| Context length (pretrain / Instruct) | 256K / 128K | company |
Paper v1 2024-11-04; the 2024-11-05 release date is not confirmed by a dated launch post I could read (HF repo created 2024-10-22 UTC, commit history truncated). Performance claims vs Llama are authors' own.
Sources
This record was checked against its sources on 6 October 2026. How we check