10M Tokens and Native Multimodality: How Llama 4 Breaks the Mold

The launch of Llama 4 marks a defining moment for Meta.

For the first time, Meta has delivered an open-weight model family—Scout, Maverick, and the massive Behemoth—built from the ground up for true multimodal intelligence.

Unlike traditional models that add multimodal features after training, Llama 4 is designed for native multimodality from the start, using a single architecture to process text, images, and video together.

Powered by a mixture-of-experts (MoE) architecture and a record 10 million token context window, Llama 4 holds longer conversations, processes more information at once, and achieves impressive results in coding, reasoning, multilingual tasks, and STEM benchmarks.

In this newsletter, we’ll explore:

• What makes Llama 4 unique

• A closer look at the Llama 4 model family: Scout, Maverick, and Behemoth

• 3 use cases that showcase Llama 4’s capabilities

• The innovations behind Llama 4’s massive training run and deployment-ready performance

Let's get started.

Meet the Llama 4 herd: Scout, Maverick, and Behemoth#

At the heart of Llama 4's release are three distinct models, each built for a different level of performance and scale:

• Llama 4 Scout: A 17 billion active parameter model featuring 16 experts, Scout is optimized for efficiency and precision. It fits on a single NVIDIA H100 GPU, yet delivers industry-leading results across vision, coding, and reasoning benchmarks.

• Llama 4 Maverick: Also a 17 billion active parameter model, but with 128 experts, Maverick scales performance even further, outperforming models like GPT-4o and Gemini 2.0 Flash on key multimodal benchmarks such as MMMU, LiveCodeBench, and GPQA Diamond (STEM reasoning). Maverick balances speed, depth, and cost-efficiency for general AI assistants and creative generation.

• Llama 4 Behemoth: Still in training, Behemoth is a 288 billion active parameter model designed to serve as the teacher for the rest of the Llama 4 family. It outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM-focused benchmarks, including MATH-500 and GPQA Diamond, and has been used to distill its capabilities into smaller, more efficient models like Maverick and Scout.

What makes Llama 4 different?#

In an increasingly competitive AI landscape, here's what sets Llama 4 apart:

• Natively multimodal: Unlike many models that layer on vision after training, Llama 4 integrates text, images, and video from the earliest stages of pretraining. This early fusion allows the model to reason seamlessly across modalities.

• Mixture-of-experts (MoE) architecture: Llama 4 uses expert routing, activating only a fraction of its total parameters for any given task. This means higher performance with lower computational cost, making large-scale intelligence more accessible.

• Record-breaking context window: With support for up to 10 million tokens, Llama 4 models can process massive inputs such as multi-document workflows, deep technical conversations, and long historical data without losing coherence.

• Benchmark leadership: Across coding, reasoning, and STEM evaluations, Llama 4 models outperform competitive offerings like Gemini 2.0 Flash and GPT-4o, while offering open weights for broader community innovation.

Training at scale: What made Llama 4 possible#

Behind Llama 4’s performance is one of the largest and most efficient open training runs. Meta trained the models using 32,000 H100 GPUs on the AI Research SuperCluster, consuming approximately 7.5 million GPU-hours to reach convergence​. Several innovations powered this scale:

• Curriculum filtering: During post-training, Meta filtered out the easiest 50% of training examples, focusing instead on medium-to-hard prompts. This curriculum-based approach significantly boosted performance in reasoning, coding, and STEM tasks.

• FP8 precision training: Llama 4 was trained in FP8 format, a lower-precision floating point technique that improved compute efficiency without sacrificing model quality, while delivering up to 390 TFLOPs/GPU in training throughput​.

• MetaP hyperparameter strategy: MetaP, a new hyperparameter scheduling technique, was introduced to optimize per-layer learning rates and initialization scales across different training scales, improving stability and final accuracy​.

Practical inference: Optimized for deployment#

Llama 4 Scout was designed not only for performance, but for efficient deployment. It supports 4-bit and 8-bit quantization, making it lightweight enough to run on a single H100 GPU, a notable achievement for a 17 B active parameter model. It delivers ~120 tokens/second throughput with ~0.4s first-token latency, enabling responsive applications at scale​.

Llama 4 in action: 3 projects you can tackle#

1. Exploring Europe through AI: Building a mini travel blog#

Let’s test Llama 4’s creative abilities by combining image generation and storytelling through Meta’s Imagine and Canvas tools.

Task: Create a short, personalized travel blog that captures the essence of iconic European locations through vivid storytelling and artistic imagery.

Llama 4’s ability to generate artistic images, match them with personalized narratives, and assemble cohesive multimedia documents will seamlessly showcase how it handles vision, text, and context. Let’s get started.

Step 1: Writing the travel blog content#

We begin by asking Llama 4 to generate the text for our travel blog. Here’s the full prompt used in Canvas:

Here’s the response generated by Llama 4:

Step 2: Generating images#

Once the blog content is ready, we generate artistic images for each location using the Imagine feature. Here is the prompt we used for the Eiffel Tower:

Here’s a response generated by Llama 4:

We can place the image within the blog’s desired location, adjust its size, edit, or delete it as needed. Similarly, we generated images for the remaining locations using the following prompts:

Finally, we added a cover art image using the “Add cover image” button, arranged the visuals inside Canvas, and completed our AI-powered travel blog. Here’s the final blog created using Llama 4:

2. From idea to game: Simulating a spinning planet#

Beyond creativity, Llama 4’s capabilities extend into structured, visual coding tasks. In this demo, we test how it can generate real-time animations, an important skill for prototyping educational simulations, scientific visualizations, and interactive content.

Task: Build a simple animated simulation of a planet orbiting a sun.

Here’s the prompt we used:

This is the code generated by Llama 4, which produces a smooth and continuous simulation of a planet revolving around a central sun.

3. Thinking through images: A story from the bazaar#

In this final task, we explore Llama 4’s ability to interpret visual input and craft meaningful narratives, combining its native image understanding with expressive language generation. We uploaded a richly detailed image of an old Moroccan bazaar, filled with colorful textiles, ornate lanterns, spice stalls, and local vendors.

Task: Write a short story inspired by the following scene.

We use the following prompt along with the image:

We can see that Llama 4 picked up on subtle visual cues and transformed them into a vivid, grounded narrative.

This file upload functionality doesn’t stop at images, and Llama 4 can also interpret documents, slides, and spreadsheets, making it a powerful tool for creative and analytical workflows.

Built for what’s next#

For years, working with AI meant trading freedom for power. Better models were locked behind paywalls and APIs, while open models often lagged. Llama 4 changes that balance.

Across creativity, code, and reasoning, Llama 4 is starting to lead. And unlike its proprietary peers, it‘s here for builders: open, accessible, and ready to be shaped into whatever comes next.

Across critical benchmarks, Llama 4 models outperform in several key areas. In coding, multilingual tasks, STEM benchmarks, and image reasoning, Llama 4‘s design choices, such as native multimodality, Mixture-of-Experts architecture, and massive context, show clear results.

Benchmarks are just one piece of the picture. The real signal is direction: larger context windows, open models handling documents, images, and logical reasoning together, not in isolation.

The launch of Llama 4 feels less a glimpse of the future of AI, i.e., multimodal, open, and powerful by design.

We’ve only just started exploring what’s possible. But if Llama 4 is any indication, the gap between closed and open is no longer about size. It’s about speed, scale, and imagination.

If you want to make the most out of Llama's features, you can build your AI skills with one of our Generative AI courses below: