Tencent’s Hy3 is a 295-billion-parameter Mixture-of-Experts AI model released on July 6, 2026, that activates only 21 billion parameters per token — making it as cheap to run as a 20B dense model while outperforming much larger systems. It ships under the Apache 2.0 license and is freely available on OpenRouter until July 21, 2026.
What Is Hy3 and Why Does the MoE Architecture Matter?
Hy3 is Tencent’s production release of what was previewed as Hunyuan 3. Its Mixture-of-Experts design uses 192 experts per layer with top-8 routing, meaning only a small subset of the total network fires on any given token. The result: hardware costs comparable to a 20B dense model, with quality that rivals models carrying three times the active parameters.
Key architecture specs verified from Tencent’s release:
- Total parameters: 295 billion
- Active parameters per token: 21 billion
- Experts: 192, top-8 routing
- Context window: 256,000 tokens
- Weight format: BF16
- License: Apache 2.0
The Apache 2.0 license is a material differentiator for enterprise teams. Unlike DeepSeek‘s more restrictive terms, Apache 2.0 permits commercial self-hosting, modification, and redistribution without legal ambiguity — the same license used by Meta’s Llama series.
Hy3 Benchmark Results: How Does It Compare?
Per Tencent’s published evaluation data and independent third-party head-to-heads, Hy3 beats DeepSeek V4 Pro on 12 of 18 shared benchmarks. It also outperforms GLM-5.1 in a 270-expert blind evaluation (scoring 2.67/4 vs GLM-5.1’s 2.51/4, per ExplainX.ai’s reported results). On scientific reasoning tasks it reaches parity with GPT-5.5.
Where it falls short: coding. On SWE-bench Verified, Hy3 scores 74.4% — strong in absolute terms, but behind GLM-5.2’s 84.2%, as reported by VentureBeat.
| Benchmark | Hy3 Score | What It Measures |
|---|---|---|
| MATH | 76.28 | Mathematical reasoning |
| GSM8K | 95.37 | Grade school math word problems |
| SWE-bench Verified | 74.4% | Real-world GitHub issue resolution |
| Terminal-Bench 2.0 | 54.4% | Agentic CLI task completion |
| Blind evaluation (270 experts) | 2.67 / 4 | Multi-domain quality vs GLM-5.1 |
Benchmark caveat: Tencent’s own published comparisons used DeepSeek-V3 and GLM-4.5 as baselines — not V4 Pro or GLM-5.2. The stronger third-party comparisons come from independent testing by CodingFleet and VentureBeat, not from Tencent’s official documentation. Worth keeping in mind before treating the headline numbers as gospel.
Where to Try Hy3 — and the Free Tier Deadline
Hy3 is available on Hugging Face, ModelScope, and Tencent Cloud TokenHub. For developers who want to experiment without deploying their own hardware, OpenRouter is offering a free ($0/token) tier through July 21, 2026 — after which standard pricing applies.
Running Hy3 self-hosted still requires significant GPU infrastructure across multiple cards despite the 21B active parameter footprint. The efficiency gains are relative to other 295B-class models, not consumer hardware.
Frequently Asked Questions
What is Tencent Hy3?
Hy3 (also called Hunyuan 3) is a 295-billion-parameter Mixture-of-Experts language model released by Tencent on July 6, 2026. It activates only 21 billion parameters per inference token, giving it the computational footprint of a 20B dense model at frontier-tier performance.
Is Hy3 open source?
Yes. Hy3 is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution. Weights are available on Hugging Face and ModelScope.
How does Hy3 compare to DeepSeek V4 Pro?
Based on third-party testing reported by CodingFleet, Hy3 beats DeepSeek V4 Pro on 12 of 18 shared benchmarks. It costs approximately half as much to run as comparable closed models in the same performance class.
Can Hy3 run on consumer hardware?
No. Despite its 21B active parameter efficiency, Hy3’s 295B total parameter count still requires significant multi-GPU infrastructure to self-host. It is not suitable for single consumer-grade GPUs.
What is the free OpenRouter tier for Hy3?
OpenRouter is offering Hy3 at $0 per token through July 21, 2026. After that date, standard usage pricing applies. Developers who want to evaluate Hy3 without deploying hardware should act before this deadline.
Last updated:
