Tencent open-sourced Hunyuan Hy3 on July 6: a 295B MoE language model that activates just 21B parameters per token, with 256K context and a fast/slow thinking switch. Apache 2.0, no geographic carve-outs. Weights are on GitHub, HuggingFace and ModelScope.
The headline claim: GLM-5.2-class results at half the size. Hy3 scores 90.4 on GPQA Diamond and wins nearly every non-coding benchmark, but loses SWE-Bench Verified 78.0 to 84.2. Tencent says agent task completion hits 90% inside its own apps — Yuanbao, CodeBuddy and ima already run on it. Hallucination rate dropped from 12.5% to 5.4% versus the previous generation.
API access and pricing
Self-host the weights, or hit Tencent Cloud’s API at about $0.15 per million input tokens and $0.59 output. OpenRouter serves it free until July 21. Best fit: long-context agent workloads and tool-calling pipelines where frontier-model pricing doesn’t pencil out.
Why it matters
Half the parameters means roughly half the serving cost at the same quality — that’s the real fight in Chinese open source right now. If your agents don’t write code all day, Hy3 just got very hard to ignore.
You Might Also Like
- Tencent Hunyuan hy3 Preview Goes Open Source 295b moe 21b Active 256k Context
- Openfang Just Dropped and its Already the Hottest Agent os on Github
- Insforge Hits 1 on Product Hunt and 3600 Github Stars is This What Agent Native Backends Look Like
- Openviking Treats ai Agent Memory Like a File System and 9k Github Stars say its Working
- Alibabas Agentscope Hits 21k Github Stars What Makes This Multi Agent Framework Different
Top AI Product 
Leave a comment