# Tencent Hunyuan Launches High-Value Large Model Hy4 Flash

- Category: Flash (快讯)
- Published: 2026-11-16T11:56:00.000Z
- Source date: 2026-11-16
- HTML: https://www.88z.ai/en/news/flash-20261116-41b55b1bd7292a58
- Markdown: https://www.88z.ai/en/news/flash-20261116-41b55b1bd7292a58.md
- Other language: https://www.88z.ai/news/flash-20261116-41b55b1bd7292a58.md

Tencent Hunyuan has released Hy4 Flash, a new high-value model designed for daily work scenarios. With 224B total parameters and only 12B activated, it supports multimodal inputs and a 1M context window. Optimized via MoE, hybrid attention, self-distillation, and reinforcement learning, it cuts output length by half while reducing costs, achieving decoding speeds over 150 tps.

## At a glance

- Hy4 Flash has 224B total parameters with 12B activated, supporting multimodal inputs and 1M context
- Optimized to halve output length and achieve decoding speeds over 150 tps

## The update

Tencent Hunyuan has officially launched Hy4 Flash, a new large model focused on high cost-performance ratio to meet daily work needs. Architecturally, it features 224B total parameters with only 12B activated, significantly boosting efficiency through sparsity. The model offers robust multimodal capabilities and supports a context window of up to 1M tokens, handling complex information processing tasks effectively. To balance performance and cost, the development team employed advanced optimization techniques including Mixture of Experts (MoE), hybrid attention mechanisms, self-distillation, and reinforcement learning. These methods not only reduce operational costs but also cut the model's output length by approximately half. Meanwhile, Hy4 Flash delivers impressive decoding speeds exceeding 150 tps, providing users with a fast and economical AI service experience.

## Sources

- [腾讯混元](https://mp.weixin.qq.com/s/5yBbf-ctYziM5r0i6Y17lA) (2026-11-16)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
