Tencent Hy3 is now available: a 295B open MoE model

Tencent released Hy3 on July 6, 2026. It’s a new open Mixture-of-Experts model with 295B total parameters, 21B active on any given task, a 256K context window, and an Apache 2.0 license.

The MoE design is the point here: each prompt only passes through part of the Hy3 network, which Tencent says cuts inference and serving costs. Hy3 is also meant to stay useful on long documents, large codebases, and multi-step work. And because it’s under Apache 2.0, you’ve got wide commercial freedom to self-host it, fine-tune it, and build global products with it, without the geographic restrictions that showed up in some earlier China-based open models.

Tencent says the July 6, 2026 release is better than the April preview after feedback from 50 internal product teams. By Tencent’s numbers, hallucinations dropped from 12.5% to 5.4%, and commonsense errors fell from 25.4% to 12.7%. Hy3 is already being used in WorkBuddy, Yuanbao, WeChat assistants, and Path of Exile: Advent, which makes it feel more substantial than a research demo built mainly to post benchmark scores.

If you want an open model for search, tool use, or agentic tasks, Hy3 looks like one to download. Coding may still be a weaker spot, since Zhipu AI’s GLM-5.2 could still have the edge there, and Tencent’s claims need independent testing either way. Still, Hy3 now joins Zhipu AI, DeepSeek, Alibaba’s Qwen, and Mistral AI in putting more pressure on closed models from OpenAI and Anthropic.

You can get Hy3 from Tencent as an open release under the Apache 2.0 license.