Today you cannot launch a new long language model, if it is not more powerful than competition. At least in the benchmarks, which do not usually serve as much. The fact is that this weekend Meta has released call 4its new multimodal AI that affects customization. Your simplest model works at home.
At the moment, the «flock» of flame 4, as the goal calls it, is composed of three AI: Call 4 Behemoth, call 4 maverick, and call 4 scout.
The last two are already available to download. Call Behemoth, the biggest model, is still trained. You have many pirate books to read …
Call 4: News and strengths
In the press release, Meta explains that this new LLM improves its predecessor when it comes to customize the multimodal AI you can use. It is native multimodal, which means that it can chat, talk, see and listen from the beginning.
Its main novelty is that Goal has changed to a mixture architecture of experts (MOE)an approach that saves resources using only the parts of a model that are necessary for a specific task, according to The Verge.
Meta Maverick It is its most versatile model, since it works in the cloud. It has 17,000 million active parameters with 128 experts. According to Mark Zuckerberg, it is the best multimodal model in its class, surpassing GPT-4O and Gemini 2.0 flash in reference tests. It achieves results similar to Deepseek V3 in reasoning and coding, with less than half of active parameters.
Meta
Call 4 Maverick offers the best performance-class ratio of its class, with an experimental version for Chat that obtained an elo of 1417 in Lmarena.
Call 4 scout is the smallest model. It presumes a context window of 10 million tokens and exceeds Gemma 3 and Gemini 2.0 Flash-Lite from Google, as well as open source Mistral 3.1, «in a wide range of widely disseminated reference tests,» according to goal.
Its main characteristic is that «it fits in a single GPU Nvidia H100», which means that Some companies can use it at home. Taking into account that these cards are worth up to 30,000 euros, it has to be a powerful company …
Finally, with its 288,000 million active parameters and 2 billion parameters in total, Call 4 Behemoth is the largest model ever created by the Facebook company. It is not yet available, but according to its creators it is more powerful than GPT-4.5. Gemini 2.0 and Claude Sonnet 3.7 in several STEM reference tests.
Many technical data and few practical examples about what it really improves with respect to the previous generation, or competition, in the press release,
Call 4, the most powerful goal generative AI has introduced today in Meta ai For the web, WhatsApp, Instagram and Messenger.
Know How we work in NoticiasVE.
Tags: Artificial Intelligence, Meta, Facebook, WhatsApp, Instagram