Tencent just made its most powerful AI model available to basically everyone. The company’s Hunyuan team released the full version of Hy3, a 295-billion-parameter Mixture-of-Experts model, under the Apache 2.0 license, removing the geographic licensing restrictions that had kept some of China’s best open-weight models out of reach for enterprises operating in the EU, UK, and South Korea.
For crypto firms that have been quietly building AI-powered trading systems, compliance tools, and on-chain analytics, the timing matters. A model with 21 billion active parameters, enterprise-grade reliability claims, and no strings attached on commercial use just entered the arena at a price point that makes OpenAI’s API look like a luxury subscription.
What Tencent actually shipped
The full Hy3 release is a significant upgrade from the preview version that launched on April 23, 2026, under the more restrictive Tencent Hy Community License. That earlier version was technically impressive but came with enough legal baggage that corporate counsel at globally operating firms would kill the deployment before engineers finished benchmarking it.
The new version solves that problem entirely. Apache 2.0 is about as permissive as open-source licenses get: use it, modify it, sell products built on it, no royalties, no geographic restrictions.
The technical specs are worth unpacking. Hy3 uses a Mixture-of-Experts architecture, meaning while the model contains 295 billion total parameters, only 21 billion are active for any given query. Tencent claims the model achieves success rates exceeding 99.99% in complex agent workflows that can involve up to 495 steps. Internal testing reportedly shows that its WorkBuddy task execution success rate can reach 90%. The model also includes a 3.8 billion Multi-Token Prediction layer designed to handle tasks requiring the model to manage multiple tokens simultaneously.
API pricing is set at 1 RMB per million input tokens and 4 RMB per million output tokens on Tencent Cloud. The model is also available on Hugging Face for self-hosting.
The licensing problem that nobody talked about
Many of the strongest models were functionally unusable for a huge portion of the global enterprise market. License terms that excluded the EU, UK, and South Korea didn’t just affect companies headquartered in those regions. Any company serving traffic into those jurisdictions, which includes most major crypto exchanges, DeFi platforms, and institutional trading desks, had to walk away.
Tencent’s switch to Apache 2.0 removes that friction entirely. The model was developed with feedback from over 50 product use cases, suggesting it was battle-tested in production environments before the public release. Performance claims include strong results on coding benchmarks, particularly SWE-Bench scores, which measure a model’s ability to resolve real-world software engineering tasks.
Why crypto firms should pay attention
A model claiming 99.99% success rates in complex multi-step workflows is exactly the kind of infrastructure that algorithmic trading desks and DeFi protocol teams need. When your trading bot executes a 50-step arbitrage strategy across multiple chains, a 95% success rate means catastrophic failures roughly once every 20 runs. A 99.99% rate changes the risk calculus entirely.
API pricing at 1 RMB per million input tokens makes it feasible to run high-frequency inference workloads that would be prohibitively expensive on GPT-4 or Claude-level pricing.
Self-hosting via Hugging Face adds another dimension. Crypto firms handling sensitive trading strategies or proprietary market data have strong incentives to keep their AI inference on-premises rather than routing queries through third-party APIs. The Apache 2.0 license makes self-hosting legally straightforward with no usage reporting requirements and no restrictions on derivative works.