Tencent opens preview access to Hy4, a 770-billion-parameter mixture-of-experts model built for enterprise coding, office, and research workflows.
The AMW Read
Incremental open-weight update from a known Chinese foundation-model player, narrowly beating domestic rivals GLM-5.3 and Kimi K3 on internal benchmarks while still trailing frontier US labs, with explicit MoE scaling detail and a deliberate open-weight distribution strategy.
Tencent opens preview access to Hy4, a 770-billion-parameter mixture-of-experts model built for enterprise coding, office, and research workflows.
Tencent released Hy4 preview as an open-weight model on August 28, using a mixture-of-experts architecture with 770 billion total parameters and 49 billion active per token, plus a 1 million-token context window. The company said it trained the model in close collaboration with engineers across software development, gaming, finance, and security to prioritize real workplace tasks over general chat ability, and tied the release to deeper integration with its CodeBuddy coding assistant and WorkBuddy collaboration platform, offering free trials to existing users of both. The model is distributed on Hugging Face, GitHub, ModelScope, Tencent Cloud's TokenHub, and OpenRouter.
In Tencent's internal blind evaluation — 163 engineers grading 203 real engineering tasks — Hy4 preview scored 2.99 out of 4.0, edging out Zhipu AI's GLM-5.3 (2.92) and Moonshot AI's Kimi K3 (2.94), with Tencent citing a large jump in coding and agentic performance over its prior Hunyuan release. That places it at the front of the open-weight cohort inside China, even as Tencent acknowledged the model still trails OpenAI's and Anthropic's frontier systems and can over-verify itself on complex tasks. The preview-first, iterate-on-feedback release pattern mirrors Tencent's prior Hunyuan open-source rollout and signals open-weight distribution remains the wedge into enterprise accounts rather than a pure benchmark play.
For builders evaluating open-weight options for coding and office-automation agents, Hy4 preview adds a third credible Chinese contender alongside Zhipu and Moonshot at comparable internal-benchmark scores, with the real differentiator being Tencent's bundled distribution through CodeBuddy and WorkBuddy rather than raw capability — worth testing against task-specific workloads before committing, given the gap to frontier US labs stays explicit even in Tencent's own framing.