Llama 5

Llama 5 is Meta's fifth-generation foundation model, released on 8 April 2026. It moves the Llama line fully onto a mixture-of-experts architecture with more than 600 billion total parameters and only a fraction of them activated per token, which is what lets Meta claim inference several times cheaper than a dense model of equivalent capability. Unlike Llama 4 it is natively multimodal across text, image, video and audio, and it was trained from the start for agentic work — tool use and multi-step planning rather than single-turn answering. Its headline number is context: five million tokens, up from 256,000 in Llama 4 and, at release, the largest window offered by any publicly available model, open or closed. Meta also describes a recursive self-improvement stage in training, in which the model refines its own reasoning traces and generates synthetic training data. On Meta's published comparisons Llama 5 gains roughly seven points on MMLU-Pro (about 87%), twelve on SWE-bench Verified (about 74%), eighteen on AIME 2025 (about 88%) and twelve on GPQA Diamond (about 84%) relative to Llama 4. Those figures are vendor-reported and have not been independently reproduced here; what is less disputed is that this is the first Llama release competitive with closed frontier models across reasoning, coding, mathematics and science at the same time. The licence is unchanged in spirit: the Llama Community Licence, not an OSI-approved open-source licence. Attribution ("Built with Llama") is still required and the 700-million-monthly-user threshold that triggers a separate agreement with Meta still applies, but Llama 5 explicitly permits using its outputs to train other models, which Llama 4's terms left ambiguous. Migration is not free: the tokenizer changed, so Llama 4 fine-tunes do not transfer and must be retrained, and the chat template and tool schema were both revised. Llama 4 remains available and has its own profile in this catalogue.

▸Videos

No videos yet.

▸Reviews