Tencent Hunyuan Launches Hy3: 295B Total Parameters, 21B Active Parameters, Performance Rivaling Trillion-Parameter Models — But The Real Highlight Is Reliability
Tencent Hunyuan has released Hy3, a 295B MoE model with 295B total parameters and 21B active parameters. Its performance matches that of trillion-parameter flagship models. With three iterations in six months, it has cut hallucination rates by half and delivers stable tool calling. It is open-sourced under Apache 2.0, with free API access available for two weeks.
Tencent Hunyuan launched Hy3 today.
From Hy2 to Hy3 Preview and now Hy3, the whole process took less than half a year.
It is a 295B total parameter MoE model, with only 21B active parameters per inference. It has 192 experts, activating the top 8 each time, and supports a 256K context window.
Those specs look decent, but what’s really impressive is its benchmark performance.
Its SWE-bench Pro score jumped from 28.6 to 62.1, HLE from 24.0 to 53.2, and BrowseComp from 28.7 to 84.2. And that’s not the most striking part —
Hy3 scores 57.9 on SWE-bench Pro, nearly double GLM 5.2's 34.7. It also leads on HLE (with tools) with a score of 53.2.
Officially, it can hold its own against flagship models with one trillion total parameters. Considering it only has 21B active parameters, this claim isn’t that far-fetched.
But the real improvement worth paying attention to is reliability.
The Tencent team conducted a blind test where 270 experts scored models based on real-world work scenarios. Hy3 scored 2.67 out of 4, beating GLM-5.1's 2.51 out of 4. Its biggest advantages are in front-end development, data storage, and CI/CD.
More concrete data:
- Hallucination rate cut from 12.5% to 5.4%
- Common sense error rate cut from 25.4% to 12.7%
- Multi-turn dialogue error rate cut from 17.4% to 7.9%
These numbers are more telling than benchmark scores. No one dares use a model that makes things up, no matter how powerful it is on paper.
Hy3 supports configurable inference modes: no_think (direct response), low, and high (deep thinking). It also received targeted optimization for tool calling stability, keeping the accuracy difference across different agent frameworks (CodeBuddy, Cline, KiloCode) within 4%.
It is open-sourced under the Apache 2.0 license, which is business-friendly. Free API access is available on OpenRouter for two weeks, through July 21.
The HuggingFace model card is already live, and deployment configurations are available for both vLLM and SGLang. It can run on just 8 H20-3e GPUs.
One netizen commented: "A 295B MoE is like a mid-sized country building a nuclear bomb in a garage." As exaggerated as that sounds, the iteration speed of Chinese large models is indeed picking up.
That said, some developers have pointed out: MoE models are prone to consistency drift in agent pipelines. A small change in context can activate different experts, leading to inconsistent output formatting. While the official labels it as "production-grade", users will still need to verify it for their own use cases.
Model releases are becoming more and more frequent, but very few can actually run stably in production. Whether Hy3 can break this trend remains to be seen as it rolls out to real-world products.
Links:
- [HuggingFace Model](https://huggingface.co/tencent/Hy3)
- [OpenRouter Free API](https://openrouter.ai/tencent/hy3:free)
- [Official Technical Report](https://hy.tencent.com/research/hy3)
发布时间: 2026-07-07 04:39