Hy3

Back to Devoured - July 07, 2026

DEVOURED

Tencent released Hy3, a 295B-parameter Mixture-of-Experts model that aims to rival flagship open-source models with fewer active parameters.

What: Hy3 is an Apache 2.0 licensed model from Tencent with 295 billion total parameters and 21 billion active parameters per forward pass. It features 256K context length and includes 3.8 billion MTP (Multi-Token Prediction) layer parameters. It is currently available on Hugging Face and free to test via OpenRouter until July 21.

Why it matters: This release continues the trend of high-performance Chinese models entering the global open-source market, aggressively challenging established leaders like Llama.

Takeaway: If you need a high-parameter model with efficient inference, test Hy3 on OpenRouter before the free access period expires on July 21.

Decoder

MTP (Multi-Token Prediction): A training approach where the model is tasked with predicting multiple subsequent tokens at once, rather than just the next single token, which can lead to better reasoning and efficiency.

Original article

tencent/Hy3. New Apache 2.0 licensed model from Tencent in China:

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. Today, we introduce Hy3, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. It also shows significant gains in utility across various products and productivity tasks.

The full-sized model is 598GB on Hugging Face, and the FP8 quantized one is 300GB. The context length is 256K.

It's available for free on OpenRouter until July 21st. I had it "Generate an SVG of a pelican riding a bicycle" there and got this:

Explore all 42 articles in today's digest