On December 5, Tencent officially launched its new self-developed large model - Huan Yuan 2.0 (Tencent HY2.0), including two versions: Think (reasoning enhanced version) and Instruct (instruction following version). It has a total of 406B parameters and activates 32B, supporting a 256K context window. The official claims that it "is in the top tier domestically" for complex reasoning tasks such as mathematics, science, and code. The model has been simultaneously launched on Tencent Cloud API and is being gradually rolled out in original applications like Yuaobao and ima.

MoE Architecture + Long Window RL, Reasoning Performance Sets New Records

- Adopting a mixture of experts (MoE) architecture, the tested single-card A100 inference speed is 40% faster than the dense structure.

- In the 2025 International Mathematical Olympiad (IMO) open questions and Harvard-MIT Math Competition, the HY2.0 Think version achieved accuracy rates of 83.1% and 81.7%, exceeding GPT-4o's score of 78.9%.

- The 256K long window solves the problem of inconsistency between training and inference through "segmented sampling correction," improving F1 by 6.8% in the 100,000 Token document QA task.

Instruction Following and Multi-turn Dialogue Optimization

- Introducing verifiable tasks and scoring-based reinforcement learning, the Instruct version scored 8.42 on Multi-Round MT-Bench, exceeding models of the same scale by 0.3 points.

- Supporting Function Call, Json Mode, and Tool Use, the official test shows a 97.2% success rate for 100 rounds of tool calls.

Accelerated Deployment: Yuaobao, ima, and Tencent Cloud Develop in Parallel

- The Yuaobao App has gradually launched the "HY2.0Think" switch, offering two scenarios: mathematical derivation and code explanation.

- The ima meeting assistant has integrated 256K long text summary, generating a 50,000 Token meeting minutes in less than 15 seconds.

- The Tencent Cloud API pricing is 0.8 yuan per million input tokens and 2 yuan per million output tokens, about 45% of GPT-4o's cost, supporting enterprise private deployment.

Open Source Plan: Model, Data, and Toolchain Launch Together

Tencent revealed that in Q1 2026, it will open-source the HY2.0Base weights, 200GB of Chinese pre-training data, and the long window RL toolchain, continuing to promote the domestic large model ecosystem.

Industry Insights: The Parameter Competition Enters the "400B+" Era