Hy3

Tencent's 295B MoE with 21B active and 256K context, released under Apache 2.0 after a preview with a territory-limited license.

In short. Hy3 is Tencent's main open language model. It is a large mixture-of-experts model that uses only 21 billion of its 295 billion parameters for each word, so it is cheaper to run than its size suggests. It can think quickly or step by step, it is aimed at agents, coding and office work, and the license allows commercial use anywhere.

Summary

Tencent officially launched Hy3 on 6 July 2026, after a preview on 23 April. It is a mixture-of-experts model with 295 billion parameters, 21 billion active per token, and a separate multi-token-prediction layer of 3.8 billion parameters. It has 80 layers, 192 experts with 8 chosen per token plus one shared expert, grouped-query attention with 64 heads and 8 key-value heads, and a context of up to 256K tokens. Tencent describes it as a hybrid fast-and-slow-thinking model. The BF16 weights are about 598 GB and an FP8 build is about 300 GB.

Tencent says Hy3 beats models of its size and matches flagship open models with two to five times as many parameters. The model card reports a blind test with 270 experts on tasks from their own work, where Hy3 scored 2.67 out of 4 against 2.51 for GLM-5.1, and internal gains such as a hallucination rate down from 12.5% to 5.4%. It also says SWE-Bench Verified accuracy varies by less than 4% across the CodeBuddy, Cline and KiloCode agent scaffolds. These are all Tencent's own evaluations, and the benchmark appendix is published as an image.

The license is Apache 2.0. That is a change from the Hy3 preview, which Tencent published under the Tencent Hy Community License, which excluded the EU, the UK and South Korea. The repository documents serving with vLLM and SGLang, fine-tuning, reinforcement-learning post-training and quantization. Tencent also offers an API through Tencent Cloud TokenHub. Hy4 preview, released in August, is a larger and newer model, still labeled a preview.

Good for

  • Agent, coding and office workflows on a multi-GPU server, with 21B active parameters per token
  • Organizations that want a large model with Apache 2.0 terms and no territory limits
  • Long-context work up to 256K tokens
  • Fine-tuning and RL experiments, since Tencent publishes the post-training instructions

Watch out for

  • The full model needs about 598 GB in BF16 or about 300 GB in FP8, so it needs a multi-GPU server
  • All benchmark and expert-evaluation results come from Tencent, and the table is an image
  • The April preview used a different, territory-limited license, so check which repository you download
  • It is text-only, with no image input

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Hy3 (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Hy3 (BF16) BF16 295B 21B active 679 GB 339 GB 204 GB Big-memory desktop
Hy3-FP8 FP8 295B 21B active 679 GB 339 GB 204 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Hy3?
Hy3 is Tencent's main open language model. It is a large mixture-of-experts model that uses only 21 billion of its 295 billion parameters for each word, so it is cheaper to run than its size suggests. It can think quickly or step by step, it is aimed at agents, coding and office work, and the license allows commercial use anywhere.
Can I use Hy3 commercially?
Yes. Hy3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Hy3 open source?
Hy3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Hy3 on my own computer?
Only on a machine with a lot of memory. The smallest variant, Hy3 (BF16), needs about 204 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Hy3 released?
Hy3 was released by Tencent on July 6, 2026.
How long a context window does Hy3 have?
Hy3 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Hy4 preview Tencent 770B 49B active 1M Apache 2.0 (Permissive) 531 GB GPU server
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
Nemotron 3 NVIDIA 31.6B – 120B 12B active 1M NVIDIA Nemotron (Permissive) 21 GB Gaming GPU