Tencent open-sources Hunyuan-A13B

Background

Tencent's Hunyuan team had already published a range of models before this release. Hunyuan-A13B was an open-source model, meaning its weights were made available for public download. The release gave outside developers a model they could download, test and run themselves.

What happened

On 27 June 2025 Tencent open-sourced Hunyuan-A13B on Hugging Face and published a technical report with it.^[1] The model is a mixture-of-experts (MoE) design with 80 billion total parameters and 13 billion active parameters, and it has a native 256K context window.^[1]

Tencent released the full-precision Instruct model together with an FP8 quantized version and a GPTQ-Int4 quantized version.^[1] Quantized versions store weights at lower numerical precision, so they need less memory to run.

Key details

The model supports both fast and slow thinking modes. Slow thinking uses chain-of-thought reasoning, in which the model writes out intermediate steps before answering.^[1]

Tencent highlighted agent benchmarks, reporting 78.3 on BFCL v3 and 54.7 on τ-Bench.^[1] BFCL v3 tests how well a model calls functions and tools, and τ-Bench tests multi-turn tool use in simulated tasks. For the Instruct model, Tencent reported 87.3 on AIME 2024.^[1]

Why it matters

The release added an open-weight mixture-of-experts model to the Hunyuan line, with its benchmark results, quantized builds and technical report published together. Developers who want to study a reasoning-capable model of this size could download the weights and the report from the same source.^[1] Hugging Face is a widely used platform for sharing model weights and related files, so the release could be reached through one place.

For readers new to the topic, the key idea is that open weights let anyone inspect the model directly, instead of relying only on the developer's own description. The benchmark figures in the technical report are still the developer's own measurements, so they are best read as a starting point for comparison. A quantized model is a smaller, cheaper-to-run copy of the same network. The trade-off is that very low precision can reduce accuracy, which is why the release included several precisions to compare.