腾讯混元Hy3发布：Agent能力和产品体验跃升

Hy3 employs a Mixture-of-Experts (MoE) architecture with 295B total parameters, but only 21B active parameters per inference, plus a 3.8B multi-token prediction (MTP) layer. The context window supports up to 256K tokens. This design aims to achieve performance close to full-parameter models at lower computational cost—Tencent officially claims Hy3 'matches flagship models 2-5 times its active parameter size.' Full-precision model weights are approximately 598GB, with an FP8 quantized version around 300GB, both open-sourced under Apache 2.0 for commercial use.

Compared to the Preview version released in April 2026, the stable version reportedly achieves significant improvements in reasoning, agent capabilities, and long-context processing. These gains are attributed to feedback collection from over 50 products and higher-quality post-training data.

Tencent released two key evaluation results. In a blind evaluation involving 270 human experts, Hy3 scored 2.67/4, higher than GLM-5.1's 2.51 (GLM-5.2 achieved similar scores in parallel tests). In internal tests, the hallucination rate dropped from 12.5% in the Preview version to 5.4%, a reduction of more than half. Additionally, Tencent claims Hy3 surpasses models of similar size on multiple benchmarks and can compete with 700B-level models such as GLM-5.1. However, these conclusions are based on official tests; independent third-party verification has not been widely conducted.

APPSO's review noted that Hy3's benchmark results are 'on par' with GLM-5.2, but did not provide specific numbers. On social media, some developers tested it and reported satisfactory performance in generating titles, tags, and SEO-friendly URLs, though these evaluations are subjective. It should be noted that the official claim of 'matching' larger models is based on specific tasks and internal testing; real-world generalization requires user judgment.

Hy3 has been released on Hugging Face, ModelScope, and GitHub, along with an FP8 quantized version. OpenRouter offers a free API until July 21, 2026, with pay-per-use pricing thereafter. Plans include supporting platforms like Cline and OpenRouter. This open strategy aims to attract community developers and accelerate ecosystem growth, contrasting with some closed-source Chinese models.

In terms of commercial deployment, Tencent has integrated Hy3 into over 50 internal products, including WorkBuddy (work assistant), CodeBuddy (coding assistant), Yuanbao (AI assistant), WeChat (social networking), ima (possibly image-related), and the Path of Exile: Adventure game assistant. This 'self-produced, self-used' model validates the model's practicality and signals Tencent's intent to rapidly convert model capabilities into product advantages in the second half of the AI race.

Despite impressive official data, some key points remain questionable. First, the claim of 'matching models 2-5 times its size' comes from Tencent itself and has not been reproduced by independent research. The blind evaluation only covered 270 experts, and the final score of 2.67 out of 4 is not a dominant lead. Second, efficient deployment of large MoE models is inherently challenging: even with only 21B active parameters out of 295B total, the memory and bandwidth requirements remain high. The full-precision model is 598GB, and the quantized version is 300GB, making it difficult for average developers to run locally.

Furthermore, Hy3's release comes at a time when the global AI wave is in its 'second half,' with intense competition among open-source models. Meta's Llama series, Mistral, and others are already mature contenders. Whether Hy3 can sustain developer interest with its Chinese language capabilities and ecosystem support remains to be seen. In the long run, its performance stability, community contributions, and Tencent's follow-up iteration plans will be key variables.