Meta has released Llama 4, its latest open-source large language model, which achieves performance within 5 percent of GPT-5 on major benchmarks while remaining freely available under a permissive license. The model comes in 70B and 405B parameter variants.
Llama 4 introduces a mixture-of-experts architecture that dramatically improves inference efficiency, making it practical to run the smaller variant on consumer-grade hardware with 48GB of VRAM. The model supports a 256,000-token context window and native tool use capabilities.
The open-source community has already begun building specialized fine-tuned versions for coding, medical, and legal applications. Meta says it has invested over $10 billion in AI infrastructure to support the Llama ecosystem this year alone.
