Tencent ships Hy3, a 295-billion-parameter open-weight reasoning model, under Apache 2.0

Tencent’s Hunyuan team officially released Hy3, a 295-billion-parameter open-weight reasoning model, on July 6, 2026, following an April preview. The weights ship under the permissive Apache 2.0 license and are available on Hugging Face and ModelScope.

The release adds another large open-weight model to a field where Chinese labs have moved aggressively on both capability and price. Hy3 is a hybrid fast-and-slow-thinking model built on a sparse Mixture-of-Experts, or MoE, architecture: it holds 295 billion total parameters but activates 21 billion at a time, routing each token to eight of 192 experts, with a 256,000-token context window.

Tencent paired the launch with low API pricing and a wide distribution push. On Tencent Cloud’s TokenHub, access runs RMB 1 ($0.15) per million input tokens and RMB 4 ($0.59) per million output tokens, with cached input at RMB 0.25 ($0.037) per million. The model is rolling out to third-party platforms including OpenRouter, Cline and Cherry Studio, and is already wired into Tencent products such as WorkBuddy, CodeBuddy and Yuanbao.

Tencent reported strong benchmark scores, including 78.0 on SWE-Bench Verified, 90.4 on GPQA Diamond and a 90 percent task-resolution rate on its own WorkBuddy enterprise platform. Those figures are vendor-supplied and have not been independently verified; benchmark results, especially on a company’s own platform, often diverge from real-world performance.

With open weights, low pricing and broad platform availability, Hy3 lands as a direct pricing and access challenge to closed frontier models. Whether developers adopt it at scale will show up in its uptake across the third-party platforms now carrying it.