Tencent open-sources the Hy3 preview

Background

Tencent's Hy Team released the Hy3 preview ahead of the final Hy3 model, which came in July 2026. A preview gives outside users an early version of a model to test before the full release.

What happened

On 23 April 2026 the Tencent Hy Team open-sourced Hy3 preview on Hugging Face, ModelScope and GitCode.^[1] Hugging Face and ModelScope are platforms where model weights are shared and downloaded, and GitCode is a code-hosting platform. Publishing on several platforms lets users reach the same model from different places.

The model is a mixture-of-experts (MoE) model with 295 billion total parameters and 21 billion activated parameters, with a 256K context.^[1] In a mixture-of-experts model, only some of the network's specialist sub-networks process each token, so the active parameter count is smaller than the total.

Key details

The preview has 192 experts with top-8 activation, which means that for each token the model uses eight of its experts. It also has one multi-token-prediction layer.^[1] Multi-token prediction trains a model to predict several upcoming tokens together, which can speed up text generation.

The preview was released under the Tencent Hy Community License Agreement.^[1] Its reasoning modes are no_think, which is the default, low and high.^[1] The base model scored 65.76 on MMLU-Pro, 95.37 on GSM8K and 76.28 on MATH with 4-shot prompting, meaning the model was shown four worked examples before each question.^[1]

Why it matters

The preview was the first open release in the Hy3 line. Its license differed from the Apache 2.0 license used for the final Hy3 model, a change readers can see by comparing the two releases.^[1] The preview also gave developers an early chance to test the architecture before the final model. Preview releases are early versions made available for testing, and details can change before the final release.