Why this matters now
On July 6, 2026, Tencent officially released Hy3 — its flagship open-weight model and the first model in the company’s newly consolidated “Tencent Hy” family. At 295B total parameters with 21B active per token, Apache 2.0 licensing, and pricing starting at $0.20 per million input tokens, Hy3 is the latest and one of the most aggressive entries in the accelerating open-weight wave.
What makes Hy3 stand out is not the benchmark scores — it’s the integration depth. Hy3 is already running in production across WorkBuddy (Tencent’s enterprise AI agent), Yuanbao (Tencent’s consumer AI assistant), WeChat customer service assistants, and even the game Path of Exile: Advent. This is not a research release hoping someone builds on it. It’s a model that shipped into real products before the press release.
Hy3 also carries Apache 2.0 licensing — full commercial freedom with no geographic restrictions — which is notable for a Chinese company release and removes the compliance friction that has shadowed some earlier open-weight models from China.
Architecture and specs
| Spec | Hy3 |
|---|---|
| Total parameters | 295B |
| Active parameters | 21B per token |
| Architecture | MoE (hybrid fast-and-slow thinking) |
| Context window | 256K tokens |
| License | Apache 2.0 |
| Model size (FP8) | ~300 GB |
| Model size (full) | ~598 GB |
| API pricing | ~$0.20/M input, ~$0.80/M output |
| Free tier | OpenRouter free until July 21 |
| Available on | Hugging Face, Tencent Cloud TokenHub, OpenRouter |
The MoE design means each query activates only 21B of the 295B parameters — roughly 7% of the network. This keeps inference costs closer to a 21B model than a 295B one, which is how Tencent hits the $0.20/M input price point.
Quality improvements from preview
Hy3 had a preview release in late April 2026. The official July 6 release incorporates feedback from 50+ internal product teams and shows measurable quality gains:
| Issue type | Preview | Official Hy3 | Change |
|---|---|---|---|
| Hallucination rate | 12.5% | 5.4% | -57% |
| Commonsense errors | 25.4% | 12.7% | -50% |
These are self-reported by Tencent and should be independently verified, but they’re consistent with the pattern of a model that improved substantially through real-product feedback rather than just synthetic benchmark tuning.
Where Hy3 beats expectations
The most interesting data point comes from Tencent’s WorkBuddy agent team. When comparing Hy3 against GLM-5.2 on document processing tasks, they reported that Hy3 completed the same workload with 47.4% fewer tokens. That’s not a benchmark score — that’s a real cost saving in a production agent. At Hy3’s pricing, 47% fewer tokens means roughly half the per-task cost on top of already-low rates.
Tencent claims Hy3 delivers “intelligence comparable to flagship models with two to five times its parameter scale.” In practical terms, they’re positioning it against models like GLM-5.2 and DeepSeek V4 on capability while undercutting both on price.
Hy3 is also a hybrid fast-and-slow thinking model — it can generate quick responses for simple queries and allocate more compute for complex reasoning, similar to the effort-level tuning in Claude Sonnet 5 and GPT-5.6 Sol.
The Apache 2.0 license matters here
Previous open-weight releases from Chinese companies have carried restricted licenses — some prohibiting commercial use in certain countries, others requiring explicit approval for high-volume deployment. Hy3’s Apache 2.0 license is clean:
- ✅ Commercial use anywhere
- ✅ Modify, distribute, sublicense
- ✅ No geographic restrictions
- ✅ Patent grants included
This removes the compliance overhead that made teams hesitate before integrating models like earlier versions of Qwen or GLM into global products. If you want to self-host Hy3 on your own infrastructure, you can — no legal review needed.
Where Hy3 fits in the market
| Dimension | Hy3 | GLM-5.2 | DeepSeek V4-Flash | Mistral Medium 3.5 |
|---|---|---|---|---|
| Active params | 21B | ~28B | 13B | ~40B |
| Context | 256K | 1M | 1M | 128K |
| License | Apache 2.0 | MIT-like | MIT | Apache 2.0 |
| Input / M tok | $0.20 | $0.73 | $0.14 | $1.50 |
| Output / M tok | $0.80 | $2.29 | $0.28 | $7.50 |
| Agent integrations | WorkBuddy, Yuanbao, WeChat | Z.ai platform | DeepSeek chat, API | Vibe, Le Chat |
Hy3 splits the difference: cheaper than Mistral and GLM-5.2 on input, more capable than V4-Flash on paper, with better licensing terms than most Chinese alternatives. The main trade-off is a smaller context window (256K vs 1M on GLM-5.2 and DeepSeek V4).
Decision framework
Use Hy3 if:
- You need an open-weight model with clean Apache 2.0 terms for a global product
- Your workloads are agentic (tool use, search, multi-step tasks) — that’s where Tencent optimized it
- You want to test it free — OpenRouter has free access through July 21
- Cost per token is your primary constraint and Mistral is too expensive
Wait or skip if:
- You need 1M+ context — Hy3 caps at 256K, and GLM-5.2 or DeepSeek V4 are better fits
- Coding is your primary use case — GLM-5.2 likely still has the edge there
- You need independent benchmark validation — Tencent’s claims haven’t been externally verified yet
- You’re already deeply integrated with Mistral or Anthropic’s ecosystem — the switching cost may not justify the price difference
The bottom line: Hy3 is the strongest signal yet that open-weight models from China are serious about global adoption. Apache 2.0 licensing, production integration at scale, and aggressive pricing make it a credible option for teams building agentic workflows — especially if they want to self-host.
Related reading
- The Agent Model Price Guide, July 2026
- Mistral’s Full-Stack Pivot: Vibe, OCR 4, and Europe’s AI Platform Bet
- GLM-5.2: 1M Context and IndexShare Sparse Attention for Coding
- Multi-Provider AI Gateways: Fallback Routing
Sources
- Tencent — Tencent Hunyuan officially releases Hy3
- Softonic — Tencent Hy3 is now available: a 295B open MoE model
- Simon Willison’s Weblog — tencent/Hy3
- GIGAZINE — Tencent has released its AI model ‘Hy3’
- Hugging Face — tencent/Hy3
- Price Per Token — Hy3 preview pricing
About the author: Charles Jasthyn De La Cueva is an Admin Officer at PSU’s Quality Assurance Office and the builder behind ParSU-EDMS / QAOSYS. He writes about AI tools, infrastructure, and practical agent deployment.
