A federal judge has ruled that Meta’s use of copyrighted books to train Llama was fair use in the Kadrey case. But the decision warns that AI could still damage the publishing market without reproducing the original works.
Meta released llama 4 scout and maverick and distinguished the experimental lmarena version from the downloadable weights. The source bounds the event: a score from one variant should not be attributed to another even when they share a family name. The story teaches how to freeze version, weights, prompt, tools and judge before comparing models.
Meta has launched Llama 3.2, a family that includes vision models and lightweight versions designed to run on devices. It has also unveiled Orion, a prototype pair of augmented-reality glasses controlled by gaze, gestures and a neural wristband.
On July 23, 2024, Meta released Llama 3.1 in 8-billion, 70-billion and 405-billion-parameter sizes. Downloadable weights increase control, but do not remove inference costs, the custom licence or the need to evaluate the whole system.
Meta has released Llama 3, with 8 billion- and 70 billion-parameter models, and rolled out its Meta AI assistant across its main apps. The company aims to make generative AI an everyday feature for billions of users.
Mixtral activates 12.9 billion of 46.7 billion parameters per token. A guide to separating compute, memory, benchmarks, variants and real openness.
French startup Mistral AI has released Mistral 7B, a 7-billion-parameter model under an open license that outperforms Llama 2 13B on several benchmarks. The company shared it with no prior announcement, via a direct download link.
Llama 2 opened weights and commercial use under a conditional license. Evaluation requires separating artifacts, rights, restrictions, costs and responsibility.
Meta released Llama 2 on 18 July 2023 with a paper carrying the word “open” in its title. The licence you must accept to download it contains two clauses the OSI’s Open Source Definition does not allow: a 700-million-user threshold, and a ban on using its outputs to improve any other model. Neither scandal nor trick — the industry norm. Here are the verbatim clauses, the contrast with the definition, and the three questions that let you read any “open” model licence in five minutes.
LLaMA showed that a smaller model can compete with a larger one, but parameters are one column. Tokens, compute, evaluation, inference and access complete the comparison.
This website uses cookies to improve the browsing experience. Cookie policy.
No results. Try different words.
