Tencent Open-Sources Hy4-preview: 770B Params, 1M Context, Strong Coding & Research Gains

智东西 · wechat · 2026-08-28

Tencent has released and open-sourced Hy4-preview, its latest flagship model: 770B total parameters, 49B activated, 1M context — a major step up from Hy3 (295B/21B/256k). Official benchmarks show it beats DeepSeek V4 Pro 0824 on coding, agent and research tasks, placing it in the top tier of open-source models. It is available on Hugging Face, GitHub, Tencent Cloud TokenHub and OpenRouter, priced at ¥6/M input and ¥18/M output tokens.

Hands-on testing by Zhidx showed notable gains in front-end aesthetics and long-horizon development: the model worked continuously for an hour to deliver a playable low-poly 3D open-world game, and built a shooter demo purely via conversation through MCP integration with Unreal Engine 5. In productivity tasks it proactively clarified requirements during procurement research and applied three-layer numeric verification in financial-report analysis, catching its own errors. Tencent reports a 31.8% end-to-end throughput gain in inference-system optimization and a 2.02x speedup in molecular dynamics simulation.

Known weaknesses include over-long thinking and excessive self-verification on complex tasks, sometimes requiring manual intervention — thinking efficiency may be the next optimization frontier.

Related event: Tencent Open-Sources Hunyuan Hy4 Preview: 770B MoE with 1M Context(28 posts)→

Original post →

More from Models

Models channel →