Mark RadarMARK RADAR
About
中 EN
Sign in
Event File AI Large Language Models (LLMs) Open-Source AI Models

Tencent Open-Sources 770-Billion-Parameter Hy4 as API Prices Jump

1 reports · First detected 2026-08-28 · Last active 2026-08-28

Tencent’s Hunyuan Hy4 preview succeeds Hy3 as the company’s open-source flagship model, targeting software engineering, office analysis, game development and scientific research. Its Mixture-of-Experts architecture activates only a fraction of its parameters for each token, seeking to balance frontier-scale capability with inference efficiency. The launch matters as Chinese developers including Tencent, Zhipu AI and Moonshot AI compete on model performance, deployment flexibility and the cost of serving increasingly complex workloads.

Tencent released Hy4 preview under the Apache 2.0 license on Aug. 28, 2026, with 770 billion total parameters, 49 billion activated per token and a 1-million-token context window. The model went live on Tencent Cloud and OpenRouter, where input pricing rose to $0.834 per million tokens from Hy3’s $0.18, a 4.6-fold increase. In Tencent’s blind test of 203 engineering tasks by 163 internal experts, Hy4 averaged 2.99 points, narrowly beating GLM 5.3 at 2.92 and Kimi K3 at 2.94.

All Coverage

1 original reports
BLOCKTEMPO.COM 2026-08-28
騰訊開源混元 Hy4 模型，7,700 億參數、價格漲超 4 倍

The Backstory

The history behind this event
Tencent Releases Open-Source Hy3 Model With Native vLLM Support2026-07-07 · 1 reports · similarity 0.80

Tencent has released Hy3 as open source under the Apache 2.0 license. The model uses a 295B-parameter mixture-of-experts (MoE) architecture, activates 21B parameters for each inference and supports a 256K context window. The release allows companies and developers to freely deploy and modify the model, reflecting Chinese technology companies’ continued investment in the open-source large-model ecosystem.

Hy3 received native support from the vLLM inference framework on its first day of release, allowing developers to deploy its MoE architecture without modifying low-level code themselves. The integration can reduce inference latency, improve decoding efficiency and lower the cost of processing long texts. The available information did not provide an exact release date or quantified performance data.

Mark Radar｜MARK RADAR
Past Issues・by MarkReadFintech

If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →

All times are in Taipei time (GMT+8)