Tencent releases Hunyuan 2.0
A 406-billion-parameter mixture-of-experts model with 32 billion active parameters and a 256,000-token context window, released in Think and Instruct variants.
- Models & capabilities
- Open weights & ecosystem
- Minor
Tencent released Hunyuan 2.0, the latest version of its flagship large language model, in two variants: a “Think” model tuned for complex reasoning and an “Instruct” model for general use. Both use a mixture-of-experts design with 406 billion total parameters and 32 billion active per token, and support a 256,000-token context window.
Tencent said the Think variant showed major gains in mathematics, science and coding over its predecessor, Hunyuan-T1, driven by changes to pretraining data and reinforcement learning, and positioned it as ranking among the top tier of Chinese domestic models on complex reasoning. No independent benchmark comparisons against non-Chinese frontier models were published alongside the release. The flagship model line was made available under a permissive open licence, while some product integrations remain commercial through Tencent Cloud.
Hunyuan 2.0 was integrated into Tencent’s own consumer AI products, Yuanbao and ima, and made accessible through Tencent Cloud’s APIs. The release continued a rapid cadence of large open or open-weight model releases from Chinese labs — alongside DeepSeek, Alibaba’s Qwen and Moonshot’s Kimi — competing on cost and openness against closed Western frontier models.