Llama
Meta's open-source family of LLMs
Start new thread

Llama 4 - A new era of natively multimodal AI innovation

by•
The Llama 4 collection of models are natively multimodal AI models that enable text and multimodal experiences. These models leverage a mixture-of-experts architecture to offer industry-leading performance in text and image understanding.

Add a comment

Replies

Best
Hunter
📌

The new herd of Llamas from Meta:


Llama 4 Scout:

•⁠ 17B x 16 experts

•⁠ Natively multi-modal

•⁠ 10M token context length

•⁠ Runs on a single GPU

•⁠ Highest performing small model


Llama 4 Maverick:

•⁠ 17B x 128 experts

•⁠ Natively multi-modal

•⁠ Beats GPT-4o and Gemini Flash 2

•⁠ Smaller and more efficient than DeepSeek, but still comparable on text, plus also multi-modal

•⁠ Runs on a single host


Llama 4 Behemoth:

•⁠ ⁠2+ trillion parameters

•⁠ ⁠Highest performing base model

•⁠ Still training!

 Just wanna leave a thread here: Llama 4 joke collectors now gather! ...



But besides we love the jokes/memes, Llama is great, people just expected 4 to be better. Fight really hard for 5 Meta!


And definitely thanks for hunting Chris!

Impressive launch for Llama 4! Curious though—how do you manage efficiency and latency challenges with the mixture-of-experts setup, especially in real-time multimodal applications?

Can't wait to try this out. We're experimenting with running models on-device for our product (desktop app) but haven't been able to get great results yet for the average laptop. Looking forward to see the reality of inference speeds for these models.

 I say Strawberry, think it is insane!

Would love to see some use cases!

What we have been waiting for!

Is the long - context ability actually viable? Judging from the evaluation, the actual performance doesn't seem to be satisfactory?

Llama 4 thinks so, more trained than the previous one

🔥 That’s one wild new herd from Meta!

Llama 4 Scout sounds like the Swiss Army knife of small models—10M context length and runs on a single GPU? That’s huge for dev accessibility. Perfect for edge devices and lightweight agents.

Llama 4 Maverick might just be the sweet spot—beats GPT-4o and Gemini Flash 2, yet compact enough to run on a single host. Multi-modal, expert routing, and smaller than DeepSeek? That’s a massive win for efficient deployments.

And then there’s Llama 4 Behemoth—the name says it all. 2+ trillion parameters?! Sounds like Meta’s going head-to-head with Gemini 1.5 Pro and GPT-5-level ambition.

⚡️ This lineup shows Meta isn’t just playing catch-up anymore—they’re coming for every tier of the LLM stack:

  • Edge → Scout

  • Mid-range agents/apps → Maverick

  • Foundation model supremacy → Behemoth

This multimodal AI is a game changer! 👀

Love how open-source models are now beating the closed source ones.
Curious if some new use cases will be opened up with 10M context length, previously even with 1M context length it's hard to direct the model what to do and usually accuracy drops.

12
Next