Is Tencent Hunyuan Hy3 open source?

Yes. Tencent released Hunyuan Hy3 under the Apache 2.0 license, with weights mirrored on Hugging Face, ModelScope, GitCode, and CNB. The code and inference recipes live in the Tencent-Hunyuan/Hy3 GitHub repository.

How is Hunyuan Hy3 different from the Hy3 preview released in April?

The July 6 release rebuilds on the April preview with cleaner post-training data and tighter training constraints. Tencent reports the hallucination rate fell from 12.5% to 5.4%, the multi-turn context-handling issue rate fell from 17.4% to 7.9%, and agent and coding scores improved 20–30% over the preview.

What hardware do I need to run Hunyuan Hy3?

Tencent's inference guide recommends serving Hunyuan Hy3 on eight GPUs via vLLM or SGLang with tensor parallelism. Only 21B parameters activate per token thanks to the 192-expert, top-8 mixture-of-experts design, so throughput is closer to a 21B dense model than a 295B one.

Where can I try Hunyuan Hy3 without downloading it?

Tencent Cloud's TokenHub exposes Hunyuan Hy3 through its large-model API. Third-party inference providers including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio are onboarding the model, with a progressive rollout listed in Tencent's launch announcement.

Tencent · 2026-07-06 · major

Tencent Hunyuan Hy3 — 295B/21B MoE open-sourced with 256K context

Tencent released Hunyuan Hy3, a 295-billion-parameter mixture-of-experts model with 21B active parameters, 256K context, and Apache 2.0 weights on Hugging Face. Tencent shares rose over 4% on the announcement.

Hugging Face social card for tencent/Hy3 model repository

Tencent open-sources its Hunyuan flagship: 295B total, 21B active, 256K context, Apache 2.0.

Key specs

Active params	21B
Context window	256K
Params (total)	295B

Quick facts

Maker	Tencent Hunyuan
Architecture	Mixture-of-experts, 192 experts, top-8 active
Total params	295B (21B active + 3.8B MTP layer)
Context window	256K tokens
License	Apache 2.0
Availability	Hugging Face, ModelScope, GitCode, CNB, Tencent Cloud TokenHub API
Third-party clients	OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, Cherry Studio

What is it?

Hunyuan Hy3 is Tencent's new flagship general-purpose model, released as open weights on Hugging Face, ModelScope, GitCode, and CNB. The full release follows the April Hy3 preview and ships a 295B mixture-of-experts model with 21B active parameters, 192 experts (top-8 activated), a 3.8B multi-token-prediction layer, and a 256K context window.

How does it work?

A hybrid fast-and-slow-thinking design switches between direct answers and step-by-step reasoning per query. Tencent's post-training added higher-quality data and stricter training constraints, cutting the hallucination rate from 12.5% to 5.4% and the multi-turn context-handling issue rate from 17.4% to 7.9%. Recommended serving is vLLM or SGLang across eight GPUs.

Why does it matter?

Tencent had lagged rival Chinese frontier labs on open weights; Hy3 catches up with an Apache 2.0 general model that Tencent claims rivals flagship open-source models 2–5x its size. Tencent-listed shares in Hong Kong closed up more than 4% on release day. Third-party integrations already include OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio.

Who is it for?

self-hosters, agent developers, Chinese-model evaluators

Frequently asked questions

Is Tencent Hunyuan Hy3 open source?: Yes. Tencent released Hunyuan Hy3 under the Apache 2.0 license, with weights mirrored on Hugging Face, ModelScope, GitCode, and CNB. The code and inference recipes live in the Tencent-Hunyuan/Hy3 GitHub repository.
How is Hunyuan Hy3 different from the Hy3 preview released in April?: The July 6 release rebuilds on the April preview with cleaner post-training data and tighter training constraints. Tencent reports the hallucination rate fell from 12.5% to 5.4%, the multi-turn context-handling issue rate fell from 17.4% to 7.9%, and agent and coding scores improved 20–30% over the preview.
What hardware do I need to run Hunyuan Hy3?: Tencent's inference guide recommends serving Hunyuan Hy3 on eight GPUs via vLLM or SGLang with tensor parallelism. Only 21B parameters activate per token thanks to the 192-expert, top-8 mixture-of-experts design, so throughput is closer to a 21B dense model than a 295B one.
Where can I try Hunyuan Hy3 without downloading it?: Tencent Cloud's TokenHub exposes Hunyuan Hy3 through its large-model API. Third-party inference providers including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio are onboarding the model, with a progressive rollout listed in Tencent's launch announcement.

Try it

huggingface.co/tencent/Hy3