Publication date: July 8, 2026
Tencent has officially released Hy3, the final version of its Hunyuan 3 foundation model, following the Hy3 preview that launched on 23 April 2026. Hy3 is an open-source Mixture-of-Experts model with 295 billion total parameters, 21 billion active parameters and a 256K context window, released under the Apache 2.0 license. According to Tencent, the model matches the intelligence of flagship models that are two to five times its parameter scale, at a fraction of the cost.
What is Tencent Hy3?
Hy3 is a hybrid fast-and-slow-thinking model built on a Mixture-of-Experts (MoE) architecture. It has 295 billion total parameters, of which 21 billion are active per token, plus a 3.8 billion parameter MTP (multi-token prediction) layer, and it supports a context length of up to 256K tokens. Developers can switch between direct responses and deep chain-of-thought reasoning through a configurable reasoning effort setting.
The official Hy3 arrives 74 days after the Hy3 preview. Tencent attributes the performance gains to strengthened reinforcement learning and improved post-training data quality and diversity. Compared with the preview, the official version also delivers greater stability and cost efficiency. Since the preview launched, average daily token consumption of the model has grown twenty-fold.
Hy3 benchmarks and technical specs
Tencent positions Hy3 as a model that performs well above its weight class. In published pre-training benchmarks, the Hy3 preview base model already competed with much larger open models such as Kimi-K2 Base (1,043B parameters), DeepSeek-V3 Base (671B) and GLM-4.5 Base (355B), while activating only 21B parameters against their 32B to 37B. It led those models on benchmarks including LiveCodeBench-v6 (34.86), GSM8K (95.37), MATH (76.28) and SuperGPQA (51.60).
The biggest gains in the official release are in coding and agent tasks. Tencent reports competitive scores on coding agent benchmarks such as SWE-bench Verified and Terminal-Bench 2.0 and on search agent benchmarks including BrowseComp and WideSearch. The company highlights productivity tasks such as software development, office productivity, financial modeling, front-end design and game production as areas of notable progress.
How is Hy3 integrated into Tencent products?
Hy3 is already deployed across Tencent products and services, including the coding assistant WorkBuddy/CodeBuddy, the consumer assistant Yuanbao, the OS-level assistant Marvis, and the knowledge tool ima. Yuanbao gained a free Agent function powered by Hy3 that can perform complex tasks and generate files in formats such as PowerPoint, Word, Excel, PDF and HTML from a natural-language description.
The model also powers AI customer service assistants on Weixin/WeChat Official Accounts and the AI assistant for Path of Exile: Advent on WeGame. Since the preview launched, the number of WorkBuddy users actively selecting the model has grown sixfold, which Tencent cites as evidence of its practical positioning.
Hy3 availability, pricing and license
Hy3 is available under the commercially friendly Apache 2.0 license and can be downloaded from Hugging Face and ModelScope from day one. Tencent will progressively roll the model out to global third-party developer platforms, including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode and Cherry Studio. The API is live on Tencent Cloud TokenHub.
Third-party hosting is already appearing: AtlasCloud lists Hy3 at 0.20 dollars per million input tokens and 0.80 dollars per million output tokens with the full 256K context window (preview endpoint). That pricing places Hy3 among the most cost-efficient options in the 200B-plus parameter class, well below flagship models such as GPT-5.5 or Claude Sonnet 5.
Conclusion
With Hy3, Tencent delivers an open-source 295B MoE model that targets flagship-level intelligence at mid-range cost, with agent capabilities as its main selling point.
The full announcement is available on Tencent’s official newsroom, and model weights can be found on Hugging Face under tencent/Hy3. Official announcement: https://www.tencent.com/en-us/articles/2202386.html
