---
title: "Tencent Releases Hy3: 295B-Parameter MoE Model with 21B Active Parameters, Apache 2.0 Licensed"
source: "Simon Willison 博客（Atom）"
source_tier: "t2"
url: "https://simonwillison.net/2026/Jul/6/hy3/#atom-everything"
permalink: "https://feed.astralor.com/items/itm_27f14f334409a8ad7749b784"
published: "2026-07-06T23:57:35.000Z"
score: 75
channel: "ai"
tags: ["Hy3","Mixture-of-Experts","Open-source LLM","Tencent"]
---

# Tencent Releases Hy3: 295B-Parameter MoE Model with 21B Active Parameters, Apache 2.0 Licensed

## AI commentary

> The highlight of this model is 'doing more with less'—only 21B active parameters challenging 70B+ models proves the efficiency of MoE. It's a boon for small and medium teams with limited budgets who want to run large models. However, independent benchmarks are lacking; I recommend testing it on OpenRouter yourself.

## At a glance

Tencent's Hy Team releases Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters. Building on feedback from 50+ products since the April preview, post-training with higher-quality data yields performance surpassing similar-sized models and rivaling flagship open-source models with 2-5x parameters. Context length 256K. Model weights 598GB (FP8 quantized: 300GB). Free on OpenRouter until July 21.

## Key points

- Scale: 295B total parameters, 21B active, 3.8B MTP layer parameters, MoE architecture.
- Performance: Outperforms similar-sized models, rivals flagship open-source models with 2-5x parameters.
- Context length: 256K tokens.
- License: Apache 2.0, free on OpenRouter until July 21.

## Full text

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: Tags: ai, generative-ai, llms, pelican-riding-a-bicycle, llm-release, ai-in-china

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_27f14f334409a8ad7749b784