If you haven't heard about "Llama 4" recently, you might be really out of date. Meta has invested heavily this time and thrown out a whole set of heavy models - Scout, Maverick, and Behemoth on the way, all of which are fierce characters.

This is not only an upgrade of the model, but also a "big discharge" to the open source AI community : they not only reconstructed the Llama series, but also embraced multimodal and hybrid expert architecture across the board, making industry peers a little uneasy.

Small but powerful: Llama 4 Scout

Don’t underestimate it because it’s the “smallest in the series”. It’s not just a small machine. It has 17B activation parameters, 16 expert units, and is incredibly fast. It can also run on a single GPU. The key is that it was designed for multimodality from birth, covering both vision and text.

What’s even more outrageous is that its context window can exceed 10 million, far exceeding most models currently on the market. This is not an ordinary “upgrade”, this is a complete makeover.

Civilian Champion: Llama 4 Maverick

Scout is a small steel cannon, and Maverick is a civilian champion. You may have heard of GPT-4o and Gemini 2.0 Flash? Now Maverick directly rubs them on the ground.

In all major evaluation lists, Maverick surpassed other products in all aspects, especially in reasoning and encoding capabilities, and its performance was close to DeepSeek v3 - the key is that it only used half of the activation parameters of the other party. Even the chat version was measured in LMArena with an ELO score of 1417, crushing a lot of "star products".

Final Boss: Llama 4 Behemoth

The first two are impressive enough, but Meta obviously has no intention of stopping there. They also revealed an ultimate boss in training - Llama 4 Behemoth.

It is said that this model has surpassed GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in STEM benchmark tests. This is still in the training stage, and it is expected to make a big splash again when it is actually launched.

Not just technology, but also attitude: fully open source

The strength of Llama 4 lies not only in its model, but also in its open attitude. Meta invited a large number of partners to work together, including Hugging Face, Together Compute, Databricks, etc.

They are not building a closed system, but opening the door wide, calling on everyone to push AI to a new stage. Llama 4 Reasoning is also on the way, and Meta's ambitions go far beyond this.

Did you know? Llama has been downloaded over 1 billion times!

Remember the news a few weeks ago? The cumulative number of Llama models downloaded worldwide has officially exceeded 1 billion! This is not an exaggeration.

Llama's success is not accidental, it is based on Meta's deep support and trust in the community, developers, and open source culture. This is exactly the biggest reason why it can quickly become popular.

What can we expect next?

The Llama 4 series is not an end, but a beginning. There are more models waiting for us in the future, including the legendary Llama 4 Reasoning.

Meta has made it clear: "This is just the beginning." For engineers, AI enthusiasts, and entrepreneurs, this is a torrent of opportunities.