Tonight’s most notable AI news is:
Tencent has officially launched the next-generation large model, Hy4 preview.
The headline figure is:
Total parameters: 770 billion.
But the more important number to understand is:
For each text request, only about 49 billion parameters are actually activated.
This highlights a key shift happening in large AI models today.
AI is no longer just about making models bigger and bigger.
It’s now focusing on:
With models this big, how can we avoid running all parameters every time?
Why not activate all 770 billion parameters at once?
Hy4 preview uses a Mixture-of-Experts (MoE) design.
Simply put, you can think of it like a very large consulting firm.
It might have:
Software engineers.
Financial analysts.
Researchers.
Documentation specialists.
Data analysts.
Different tasks call for different experts.
If the client only needs a code fix, you don’t gather all thousands of employees into one meeting.
The system determines which “experts” to activate, turning on only part of the model.
This is exactly how Hy4 preview works.
The entire model is huge, but each computation only involves a small fraction.
This is why the “total parameter” count alone is becoming less useful to gauge an AI’s practical complexity.
Why does this matter to everyday users?
Because once AI models are used daily, three factors matter beyond raw capability:
Speed.
Cost.
Computational resources.
A model that answers very well but needs an entire data center for every query can’t scale widely.
So the next competition is not just:
Whose model is the biggest?
It becomes:
Who can maintain a vast knowledge base but activate only what’s needed each time?
It’s like a company not stuffing all employees into every meeting but letting the right people handle the right tasks.
Hy4 preview is more than coding
Tencent’s positioning of Hy4 preview isn’t simply as a coding model.
Officially, its application scenarios cover:
Software engineering.
Office work.
Scientific research.
And other tasks needing long, deep comprehension of large information volumes.
Reuters also notes it is positioned for coding, research, and financial analysis.
Additionally, its context window—the amount of content the AI can reference at once—exceeds 1 million tokens.
You can roughly think of tokens as the AI’s basic reading units.
The key is not the number “1 million” itself.
The real significance is:
AI is being designed to handle longer projects, not just answer a small snippet of text.
But this is still just a preview
This shouldn’t be overlooked.
The name includes "preview" for a reason.
Reuters reports Tencent admits this early version may spend more time than necessary reasoning on complex problems and sometimes double-check its answers.
So don’t jump to the conclusion that:
This is already the best AI model just because of 770 billion parameters.
Tencent’s internal benchmarks show where they think the model stands, but internal testing isn’t the same as full, independent verification.
The real test will be:
How external developers use it.
Its performance on various tasks.
Speed.
Cost.
Stability.
And whether its ultra-long context abilities get used in daily workflows.
Why did Tencent open source it?
This might be even more important.
Tencent not only released Hy4 preview but also opened it for external developers.
Simultaneously, they are integrating Hy4 preview into products like CodeBuddy, WorkBuddy, and their API ecosystem.
This shows large AI companies are now competing on three fronts simultaneously:
Model capabilities.
Developer adoption.
Real workplace integration.
A model can top benchmarks but without developer uptake, product connections, or enterprise workflows, its commercial value remains limited.
Conversely, an open source model widely adopted by tools, platforms, and companies can quickly build its own ecosystem.
Yesterday Hugging Face, today Tencent Hy4—they represent the same trend
Yesterday’s report covered Nvidia reportedly preparing to acquire Hugging Face.
Today, Tencent drops a new large model into the open source ecosystem.
Putting these together shows AI competition is no longer just rankings of models.
The real battleground covers the entire chain:
Who builds the model.
Who provides computing power.
Who helps developers find the model.
Who integrates models into business workflows.
And ultimately, who controls the daily AI usage entry points.
So what’s most notable about Hy4 preview isn’t the big 770 billion number.
It’s that large AI models are evolving into “expert organizations”:
Big capabilities,
Activate only what’s needed per task,
And allow outside developers to use them.
The next AI race may not be who can build the largest model,
But who can turn a big model into a truly usable, integratable, and everyday working tool.
Today, let’s improve a bit with AI.
Learn one AI trick every day.
Save a little time every day.
Increase your ability a little every day.
SasaDaily grows with you.
