Llama

Llama is Meta’s family of large language models. Meta does not only serve it behind a hosted API: the weights can be downloaded and run on your own hardware, so inputs and outputs never leave the machine.

The family spans several sizes, from small edge models to server-class ones, and the licence it is released under puts conditions on redistribution and commercial use.

What is in the family?

The current generation is Llama 4, two models that the model card describes as natively multimodal, both built on a mixture-of-experts architecture:

• Llama 4 Scout: 17 billion active parameters with 16 experts, 109 billion total, 10 million token context window.
• Llama 4 Maverick: 17 billion active parameters with 128 experts, 400 billion total, 1 million token context window.

The model card lists 12 supported languages, Hungarian not among them, and pretraining covered a wider set that Meta allows fine-tuning to extend under the licence and the acceptable use policy. Earlier generations remain available, including Llama 3.2 1B and 3B for edge use, and the small ones fit on a single machine.

Access is not anonymous: the llama.com download page asks for a legal name, date of birth and organisation details, and the licence has to be accepted. The models also reach official Hugging Face repositories and Meta’s hosting partners, whose own terms may apply.

Hardware and quantisation

Meta’s FAQ states that requirements depend on the model and on latency, throughput and cost. Larger models are typically split across several inference chips with tensor parallelism. They run on GPUs, x86 and ARM CPUs, TPUs, NPUs and accelerators, and smaller ones also on system-on-chip platforms. The Llama 4 model card adds specifics:

• Scout is released as BF16 weights and, per Meta, fits on a single H100 GPU with on-the-fly int4 quantisation.
• Maverick ships as both BF16 and FP8 quantised weights, and the FP8 version fits on a single H100 DGX host.

The price is documented: Meta provides quantised checkpoints for deployment flexibility, but reports that all evaluations ran on bf16 models. The trade-off is not free, so we measure it on our own task.

What the licence says

Llama 4 is released under the Llama 4 Community License Agreement of April 5, 2025, which the model card calls a custom commercial license, not a standard open source licence. The grant is non-exclusive, worldwide, non-transferable and royalty-free, and covers use, reproduction, distribution and derivative works.

• On redistribution you include a copy of the agreement and prominently display “Built with Llama” on a related site, interface or documentation.
• Derivative models trained on Llama outputs must begin with “Llama” in their name.
• Attribution: distributed copies keep the Notice file with the copyright line.
• Above 700 million monthly active users a separate licence must be requested from Meta.
• Acceptable use policy: part of the agreement, listing prohibited uses such as operating critical infrastructure.
• EU restriction: the policy of Llama 3.2, 3.3 and 4 restricts EU based companies and individuals on the multimodal models.

The agreement disclaims warranties, and litigation against Meta terminates the rights.

Where we use it

We run local models with LM Studio and expose them over OpenAI-compatible endpoints on our own machines, so data stays on the host and Hermes Agent calls them like a remote service. Docker provides the frame where several processes have to be wired together, and our Hungarian translation jobs also run on a local model.

Further reading

• Llama 4 Community License Agreement - the full licence text.
• Llama 4 model card - architecture, quantisation and supported languages.
• Llama FAQ - hardware, licensing and the EU restriction.