Meta introduces Llama 4 with new advanced multi-modal LLMs
Meta has launched two new AI models, Llama 4 Scout and Llama 4 Maverick, as part of its Llama 4 suite, with plans to release additional models, Behemoth and Reasoning, soon. Llama 4 Maverick is tailored for general assistant and chat applications, excelling in image and text comprehension. In contrast, Llama 4 Scout, a smaller model, specializes in multi-document summarization, personalized task parsing, and reasoning over extensive codebases.
Scout features 17 billion active parameters with 16 experts, offering fast, natively multimodal performance and a 10 million token context length on a single GPU. Maverick, with the same parameter count but 128 experts, reportedly surpasses models like GPT-4o and Gemini 2.0 in coding, reasoning, multilingual tasks, and image benchmarks. Meta also announced the upcoming Llama 4 Behemoth, boasting 288 billion active parameters and touted as one of the world's most advanced LLMs.
Meta's newest Llama 4 models are already available through third-party services like Amazon Web Services, Azure AI Foundry, and Cloudflare's Workers AI. They can also be downloaded directly from the Llama website and Hugging Face, and have been integrated into Meta AI platforms such as WhatsApp, Messenger, and Instagram DMs. However, Meta AI’s integration into WhatsApp has faced criticism for not allowing users to remove or disable the chatbot, leading some to switch to WhatsApp Business or other messaging platforms without AI features like Signal or Session.
- Large Language Model (LLM) Tool
- Free • Open Source
Related news
External links
- The Llama 4 herd: The beginning of a new era of natively multimodal AI innovationMeta AI • Official source
- Meta launches new Llama 4 AI for all your apps, but it still feels limited compared to what ChatGPT and Gemini can doTechRadar • Official source
- Meta releases Llama 4, a new crop of flagship AI modelsTechCrunch
- Meta Releases Llama 4 AI Models; Beats GPT-4o and Grok 3 in LMArenaBeebom
- Meta just launched Llama 4 — here's why ChatGPT, Gemini and Claude should be worriedTom's Guide
- Meta's Llama 4 models now available on AWSAmazon Web Services
- Introducing the Llama 4 herd in Azure AI Foundry and Azure Databricks Microsoft Azure Blog
- Meta’s Llama 4 is now available on Workers AIThe Cloudflare Blog
Comments
On performance/relevance side, we still need more hindsight to correctly appreciate what Meta has done. But "smarter" (more coherent) and efficient LLMs is still good new. What is worrying is the reducing of safeguards. Theses last months, Facebook/Meta/Mark has estimated (following current US policies) that biais are worse on one side (the one they used to be) that the other one, and allowed many more things to be said on their platform (which is yet another subject). But as constantly repeated, LLM have no ideas of what they're talking about, and giving how much some very strange theories are plaguing the web (but no citing any kind of proof, just other strongly biased websites that tell the some things), LLMs need very strong safeguards to avoid spreading complete nonsense (and being irrelevant, if not dangerous). And Meta is explicitly trying to align Llama on the same irrelevance scale than Grok, because the current biased politic climate seems a better goal than objective knowledge (i.e. what most benchmarks are trying to test). And yes, tool needs to be neutral, because that what tools are for.
Tools like that cannot be neutral. They could if copyright didn't existed and that Open Source model would 100% open source, from training data, tweaking to all the training methods. As of today no good LLM is really open source (mainly because of copyright law but also competition)