Tencent has open sourced Hunyuan Hy3, a 295 billion parameter language model, releasing the full weights under the Apache 2.0 license on Hugging Face, ModelScope, and GitCode. The launch follows a preview version that Tencent put out in April, and it now includes both the standard build and an FP8 variant that is cheaper to run. Apache 2.0 is a permissive license that allows commercial use, which matters as much as the raw capability, because it means companies can build on Hy3 without the usage restrictions that some other model licenses impose.
The number that defines the model is not really the 295 billion total parameters, it is the roughly 21 billion that are active at any one time. Hy3 is a mixture of experts model, an architecture that stores a very large set of specialized subnetworks but only activates a small fraction of them for each token. The effect is that the model can carry the knowledge of something much larger while costing far less to run, closer to a 21 billion parameter model than a 295 billion one. It ships with a 256K token context window, which is long enough to hold large codebases or document sets in a single pass, and the FP8 build lowers the memory and compute needed to deploy it further.
On capability, Tencent leans on a blind evaluation rather than the usual leaderboard of automated benchmarks. It says 270 experts scored competing models on tasks taken from their own day to day work, and Hy3 landed at 2.67 out of 4, ahead of GLM-5.1 at 2.51. The company reports particular strength in reasoning, agentic workflows, long context tasks, coding, office productivity, financial modeling, frontend design, and game development, and it claims the model rivals flagship open systems that run several times as many active parameters. The blind, work based format is a more meaningful signal than a synthetic benchmark, though it is still the vendor reporting its own result.
It is worth being precise about what the claim is and is not. Beating GLM-5.1 in a blind test and leading its size class is a real and specific result. Rivaling flagships with two to five times the active parameters is a strong efficiency argument. Neither of those is the same as topping the closed frontier models from the largest labs, and Tencent is not really claiming that. The honest read is that Hy3 is a capable, efficient, and fully open model that is very good for its active size, not a declaration that the open world has caught the very top of the closed one.
The reason it matters beyond one release is that it keeps extending a pattern that has run all year, in which large Chinese technology companies and labs keep shipping capable models under genuinely open, commercially friendly licenses. Hy3 sits alongside a steady run of open weight releases in 2026, and it reinforces two of the year's clearest trends at once, the move toward mixture of experts designs that keep active compute low, and the push toward agentic capability as the thing these models are tuned to do. For anyone building on open models, a 295 billion parameter system that runs like a 21 billion one, is strong at agents and long context, and carries an Apache 2.0 license, is a genuinely useful new option rather than just another entry on a list.