Yuanyou Released a new collection of AI modelsCamel 4, in its Camel family – on a Saturday, no less.
There are four new models in total: the Llama 4 Scout, the Llama 4 Maverick and the Llama 4 Behemoth. All of this is trained in “a large amount of unmarked text, images and video data” to give them a “wide visual understanding,” Meta said.
The success of the open model of China’s artificial intelligence laboratory DeepSeekIt reportedly performed better or better than Meta’s previous flagship llama model, reportedly bringing Llama development into speeding. Meta is said to have a stir-fry room to decipher how DeepSeek reduces the cost of running and deploying models R1 and V3.
Scouts and Mavericks are publicly available on llama.com and are publicly used from Meta’s partners, including AI Dev Platform, while Bememoth is still in training. Meta AI, an AI-powered assistant to WhatsApp, Messenger and Instagram, has been updated to use Llama 4 in 40 countries. The English is temporarily in English, and the multi-mode function is limited to the United States.
Some developers may question the Llama 4 license.
Users and companies that use or distribute models are prohibited from “residence” or have “primary business locations” in the EU, which may be the result of governance requirements under the AI and data privacy laws in the region. (In the past, Meta condemned these laws to be too heavy.) In addition, like previous Llama distributions, companies with more than 700 million active users must apply for a special license from Meta, which Meta can approve or reject it in its entirety.
“These Llama 4 models mark the beginning of a new era for the Llama ecosystem,” Mehta wrote in a blog post. “This is just the beginning of the Llama 4 series.”
Meta said Llama 4 is its first model to use a mix of experts (MOE) architectures, which is more effective for training and answering queries. The MOE architecture basically breaks down data processing tasks into subtasks, and then delegates them to smaller professional “experts” models.
For example, the Mavericks have a total parameter of 400 billion, but only 17 billion Positive 128 “Expert” parameters. (The parameters roughly correspond to the problem-solving skills of the model.) Scout has 17 billion active parameters, 16 experts and 1009 billion parameters.
According to internal testing by Meta, the company said Maverick is best suited for “ordinary assistant and chat” use cases such as Creative Writate, surpassing OpenAI’s GPT-4O and Google’s Gemini 2.0, involving certain encoding, reasoning, multilingual, long posts and image benchmarks. However, Maverick doesn’t quite measure more recent models, such as Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 sonnet, and OpenAI’s GPT-4.5.
The advantage of scouts is in tasks such as document summary and large codebase reasoning. Uniquely, it has a large context window: 10 million tokens. (“Token” represents bits of the original text – for example, the word “wonderful” is divided into “fan”, “tas” and “tic”.) In plain English, Scouts can take images and up to millions of words, making it possible to process and use very large documents.
According to Meta, Scouts can run on a single NVIDIA H100 GPU, while Maverick requires an NVIDIA H100 DGX system.
Meta’s unreleased behemoth even requires more powerful hardware. According to the company, Behemoth has 288 billion active parameters, 16 experts and nearly 2 trillion parameters. Meta’s internal benchmarks outperformed GPT-4.5, Claude 3.7 SONNET, and GEMINI 2.0 PRO (but not 2.5 Pro), and these evaluations measure STEM skills such as mathematical problem solving.
It is worth noting that neither the Llama 4 model is the correct “inference” model along Openai’s O1 and O3 Mini. Inference models fact check their answers and generally answer questions more reliably, but therefore take longer to provide answers than traditional “non-disputed” models.
Interestingly, Meta said it adjusted all Camel 4 models to refuse to answer “controversial” questions. According to the company, Llamas 4 responded to the political and social topic of “debate” that the previous Rama model would not. Additionally, the company said the Llama 4 is “more balanced”, which prompts it can’t be entertained.
“[Y]You can count on [Lllama 4] A Metal spokesperson told TechCrunch to provide a useful factual response.[W]E’RE continues to make Llama more responsive so that it can answer more questions that can answer a variety of different opinions […] And it does not support other views. ”
These adjustments are accused by White House allies of AI politics of suffocation.
Many close confidants of many President Donald Trump, including Elon Musk, Cryptocurrency (Ai Czar) and AI “Tsar” David Sacks, claiming many AI chatbots Review conservative perspectives. Sacks have historically Pick it out Openai’s Chatgpt is especially “programmed to wake up” and feels unreal to politically sensitive subjects.
In fact, AI bias is a thorny technical problem. Musk’s own AI company Xai has struggle Create a chatbot that doesn’t recognize other political views.
This has not stopped companies including Openai Adjustment Their AI models answer more questions than before, especially on controversial political topics.
