Meta Unveils Llama 4

Meta has raised the bar for AI performance with its new Llama 4 Scout and Llama 4 Maverick models — smaller, faster, and more capable than many industry giants. These models are not just smaller and faster; they’re also more capable than many of the big names in the industry. With open weights now available for download, they promise to deliver outstanding performance across text, image, and reasoning tasks — all while being super efficient.

A Leap in Multimodal AI

The star of the show is Llama 4 Maverick, a 17B-parameter model with 128 experts, designed to outperform rivals like GPT-4o and Gemini 2.0 Flash in coding, reasoning, and image understanding — despite using half the active parameters of competitors like DeepSeek v3. It also boasts a best-in-class performance-to-cost ratio, with a chat version scoring 1417 ELO on the LMArena benchmark.

Meanwhile, Llama 4 Scout — another 17B-parameter model, but with 16 experts — sets a new industry standard with a staggering 10 million-token context window, ideal for parsing massive documents or codebases. Both models were distilled from Meta’s upcoming Llama 4 Behemoth, a 288B-parameter “teacher model” still in training, which already outperforms GPT-4.5 and Claude Sonnet 3.7 on STEM benchmarks.

Why It Matters

Meta’s commitment to native multimodality means these models can effortlessly combine text and visual inputs, leading to more engaging AI interactions. Thanks to early fusion training and an upgraded vision encoder, they can analyze up to 48 images at once, with precise “image grounding” that connects prompts to specific areas in a picture.

The company also highlighted its dedication to openness by releasing the models on platforms like Hugging Face and llama.com, along with a focus on safety. Tools like Llama Guard are in place for content moderation, and Llama 4 is designed to refuse fewer controversial topics compared to its predecessor.

What’s Next?

Developers can integrate Llama 4 into apps today, while Meta AI powered by these models rolls out in WhatsApp, Messenger, and Instagram. Meta teased more updates at LlamaCon on April 29, hinting at Behemoth’s future release.

With unmatched efficiency and multimodal prowess, Llama 4 isn’t just an upgrade — it’s Meta’s bid to redefine the AI landscape.