Overview
Hunyuan Hy3 is a 295B-parameter Mixture-of-Experts language model developed by the Tencent Hy Team (the rebranded Hunyuan group) and released on 6 July 2026. It activates 21 billion parameters per token across 192 experts (top-8 activated), adds a 3.8B Multi-Token Prediction layer, uses grouped-query attention with 8 KV heads, and serves a 256K context window. Weights are published on Hugging Face as tencent/Hy3 (BF16) and tencent/Hy3-FP8 under the Apache 2.0 license, making the model free to download, fine-tune, and use commercially.
Hy3 follows the Hy3 Preview released in late April 2026: Tencent gathered feedback from more than 50 products and scaled up post-training with higher-quality data, and the Hy team positions the final release as an agentic, coding-focused model that rivals flagship open-weight systems two to five times its size. The model exposes a switchable reasoning-effort mode (`no_think`, `low`, `high`) via `chat_template_kwargs.reasoning_effort`, letting callers trade latency for depth on the same weights.
Alongside the raw benchmark work, Tencent reports substantial product-reliability gains from Hy3 Preview: hallucination rate on internal real-world evaluations dropped from 12.5% to 5.4%, commonsense error rates fell from 25.4% to 12.7%, and multi-turn issue rates on internal comprehensive tests dropped from 17.4% to 7.9%. A blind evaluation with 270 domain experts scored Hy3 at 2.67/4 versus GLM-5.1 at 2.51/4, with the strongest gains in frontend development, data & storage, and CI/CD work. On SWE-Bench Verified, cross-scaffolding accuracy variance across CodeBuddy, Cline, and KiloCode stays within 4%.
| Released | 2026-07-06 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 295B total · 21B active |
| Context | 256K |
| Architecture | Mixture-of-Experts (192 experts, top-8 activated) |
| Modalities | Text |
| Status | Generally available |
Strengths
- Apache 2.0 open weights at 295B total parameters (21B active) — free for commercial and self-hosted use
- 256K-token context window with grouped-query attention and Multi-Token Prediction speculative decoding
- Switchable reasoning-effort setting (no_think / low / high) on the same weights
- Consistent behaviour across CodeBuddy, Cline, and KiloCode agent scaffoldings (SWE-Bench Verified variance within 4%)
- Distributed on Hugging Face, ModelScope, GitCode, CNB and Tencent Cloud TokenHub, with an FP8 variant for lower-VRAM serving
Best for
- Self-hosted agentic coding, tool use and terminal automation via vLLM or SGLang recipes
- Long-document and long-codebase analysis that needs the full 256K-token window
- On-prem enterprise deployments where an Apache-2.0 licence is a requirement
- Reasoning-heavy research and evaluation workloads that need a switchable effort mode
How to access
| Provider | Model ID |
|---|---|
| Tencent Cloud TokenHub ↗ | hy3 |
| Hugging Face (self-hosted via vLLM / SGLang) ↗ | tencent/Hy3 |
FAQ
Is Hunyuan Hy3 open source?
Yes. Tencent published Hy3 (BF16) and Hy3-FP8 on Hugging Face, ModelScope, GitCode, and CNB under the Apache 2.0 license, so the weights can be downloaded, fine-tuned, and used commercially, including for self-hosted and on-premise deployment. Hy3-preview, released in April 2026, used Tencent's Hunyuan community license; the July 6, 2026 Hy3 release moved to Apache 2.0.
How large is Hy3 and how much context does it handle?
Hy3 is a Mixture-of-Experts model with 295 billion total parameters, of which 21 billion are activated per token across 192 experts (top-8 activated). It ships an additional 3.8 billion parameters in a Multi-Token Prediction layer for speculative decoding. Context length is 256K tokens, and Tencent publishes both a BF16 checkpoint and an FP8 quantized variant for lower-VRAM serving.
What are Hy3's reasoning modes?
Hy3 exposes a single switchable reasoning-effort setting through `chat_template_kwargs.reasoning_effort` on the OpenAI-compatible API: `no_think` for a direct answer (the default), `low` for a shorter chain of thought, and `high` for deep chain-of-thought on hard math, coding, and reasoning tasks. The same weights serve every mode.
How do I run Hy3 locally?
Tencent recommends serving Hy3 on 8 GPUs with large memory (H20-3e-class or better), using vLLM or SGLang. Both have dedicated recipes: `vllm serve tencent/Hy3 --tensor-parallel-size 8 --speculative-config.method mtp --tool-call-parser hy_v3 --reasoning-parser hy_v3` for vLLM, or `python -m sglang.launch_server --model tencent/Hy3 --tp-size 8 --tool-call-parser hunyuan --reasoning-parser hunyuan --speculative-algorithm EAGLE` for SGLang.