Tencent · 2026-04-23 · major
Tencent Hy3-Preview — 295B Open-Source MoE with 74.4% SWE-bench
Tencent open-sources Hy3-preview, a 295B MoE model (21B active params) scoring 74.4% on SWE-bench Verified and 87.2% on GPQA Diamond. Led by ex-OpenAI researcher Yao Shunyu; ships day-1 with 256K context and adjustable reasoning modes.
Tencent's first open-source flagship under ex-OpenAI researcher Yao Shunyu — 74.4% SWE-bench in a 295B MoE.
What is it?
Hy3-preview is a 295B-parameter Mixture-of-Experts model from Tencent's Hunyuan team, released as open weights on HuggingFace. It activates 21B parameters per token and achieves 74.4% on SWE-bench Verified. It is the first model shipped under Yao Shunyu, formerly of OpenAI, who joined Tencent to rebuild their foundational AI research infrastructure from scratch.
How does it work?
MoE architecture with 192 experts, top-8 activated per token. The model supports three inference-time reasoning modes (no_think / low / high) to trade compute for analytical depth. Context window is 256K tokens at BF16 precision.
Why does it matter?
74.4% on SWE-bench Verified is among the highest scores for any open-weights model at this scale. It ships day-1 into Tencent's consumer app Yuanbao and coding assistant CodeBuddy, and is accessible via OpenRouter's free tier today. Practitioners get a capable coding and reasoning model they can run or call without a frontier-lab waitlist.
Who is it for?
ML engineers and self-hosters who need a powerful coding and reasoning model with open weights.
Try it
openrouter.ai/tencent/hy3-preview