Tencent open-sources Hy4, a 770B-parameter AI model

Tencent has released and open-sourced Hy4 preview, a large language model with 770 billion total parameters, making it one of the largest openly released models to date.
The figure that matters most for running it is smaller. Hy4 uses a mixture-of-experts design, in which only a fraction of the network handles any given token. About 49 billion parameters are active per token, spread across a backbone of 78 layers, 77 of which use the mixture-of-experts structure with 256 routed experts and one shared expert. That keeps the cost of serving the model far below what a 770-billion-parameter figure would suggest.
The context window exceeds 1 million tokens. Tencent positions the model for productivity work: coding, document preparation, data analysis, game development and scientific research. The company says training used data created with its own domain experts across software engineering, gaming, finance and security.
On performance, the numbers released so far come from Tencent itself and should be read as such. In an internal blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview averaged 2.99 out of 4.00, which the company says places it slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. Those margins are narrow, and no independent benchmark of the model has been published yet. Tencent also reported a 31.8 percent throughput gain in its own systems.
The model is on Hugging Face and is reachable through Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima, with API access via Tencent Cloud TokenHub and OpenRouter. API pricing is listed at $0.834 per million input tokens and $2.501 per million output tokens, with cache hits at $0.042 per million.
Tencent describes this as a preview and says the next batch of models in the Hy4 series will follow. The company has not detailed the specific licence terms attached to the release.
Disclosure: NewUJ’s editorial process uses Anthropic’s Claude models, and Anthropic competes with Tencent in AI models.
Related
Claude Finds CRISPR-Like Enzyme System in 21-Hour DNA Search
Anthropic Cuts Opus Costs 40%, OpenAI Answers in an Hour
StepFun's 600B Step 5 Scores 44 at $1 per Million Tokens
Claude Leads 26% of Anthropic's AI R&D, Up From 1% in March
Universal, Sony Sue Suno Again Over 60,202 Recordings
Alibaba Open-Sources CT AI That Flags 146 Conditions
Claude Sped Up 30+ Biology Models 4x, $1M Protein Contest Opens
King Charles to AI Leaders: Control 'Before It Is All Too Late'
Trending now
- US-China Trade Truce Extended to Jan. 10 as Xi Visits
- Amoeba Breeds at 63°C, Past the 60°C Limit for Complex Life
- 996 Zyxel Switches Looted in 48 Countries; Patch Due Sept. 24
- Oracle Invokes Force Majeure on 2.45GW Stargate Data Center
- Diller Drops $18B MGM Bid; Stock Falls 9.5% to February Levels
- Taylor Swift Adds 4 Songs to ‘Showgirl’ in Sept. 25 Encore
- Attacks on WordPress RCE Flaw Rose Tenfold After Sept. 22 Fix
- Rivian Recalls 98,828 EVs Over Rearview Camera Fault
Comments
No comments yet. Be the first.