Tencent released Hunyuan 2.0, a large language model with 406 billion total parameters, on December 5. This update targets advancements in mathematics, coding, and complex reasoning through a Mixture-of-Experts architecture and refined training methods.
The model comes in two variants: Think and Instruct. Its architecture activates 32 billion parameters during each inference, enabling efficient processing. It supports a context window of 256,000 tokens, allowing handling of extensive inputs without truncation.
Tencent positions HY 2.0 Think as ranking among the “top tier domestically” for complex reasoning tasks. This version surpasses its predecessor, Hunyuan-T1-20250822, in multiple evaluation areas. On the IMO-AnswerBench mathematics benchmark, HY 2.0 Think recorded a score of 73.4, reflecting strong problem-solving capabilities in mathematical domains.
In software engineering assessments, performance improved markedly on the SWE-bench Verified benchmark, rising from 6.0 for the prior model to 53.0. This enhancement demonstrates better accuracy in generating and debugging code for real-world programming challenges.
Tencent HY 2.0 is officially released. We are rolling out a major performance upgrade to our foundation model, now available via Tencent Cloud API.
Built on a Mixture-of-Experts (MoE) architecture (406B total, 32B active parameters) and featuring a 256K context window, HY 2.0… pic.twitter.com/zmb2zLQTEz
— Hunyuan (@TencentHunyuan) December 5, 2025
Tencent credits these advancements to improvements in pre-training data quality and a dual-stage reinforcement learning approach. This strategy integrates RLVR, or Reinforcement Learning with Verifiable Rewards, which uses objective metrics for training, and RLHF, or Reinforcement Learning from Human Feedback, incorporating human evaluations to refine outputs.
HY 2.0 Think applies length-penalty strategies to prevent excessive verbosity in responses, resulting in what Tencent calls “industry-leading” computational efficiency per token. This design optimizes resource use during generation, reducing processing time and costs.
Coding and agent functionalities also advanced, with the Tau2-Bench score climbing from 17.1 to 72.4. These metrics evaluate autonomous task execution and code-related interactions.
Integration extends to Tencent’s consumer applications, such as Yuanbao and ima, where the model enhances user interactions. Developers can access it via Tencent Cloud’s API platform for custom implementations. Tencent plans to open source related technologies and models for community use.
