Meta releases Llama 3.3 70B, matching much larger models at lower cost
Originally announced by Meta . We link the primary source so you can read it for yourself.
Meta has released Llama 3.3 70B Instruct, a 70-billion-parameter open-weight model that the company positions as delivering performance close to its much larger 405B model at a fraction of the hardware cost.
The news
Llama 3.3 70B is an instruction-tuned text model with a 128k context window, released under the Llama 3.3 Community License. Meta reports a score of 86.0 on MMLU for the model (model card, December 2024). As with any figure from a releasing organisation, treat that as a claim to verify against independent evaluations rather than a settled result.
The weights are available on Hugging Face, and the model works across the major inference engines. In GGUF form it runs through llama.cpp and Ollama; an MLX build covers Apple Silicon.
What it means
The interest here is the size-to-capability ratio. A 70B model is within reach of a well-specified desktop or a Mac with enough unified memory, whereas a 405B model is not. At a 4-bit quantisation the model needs roughly 43GB of memory, which puts it on a single high-memory professional card, a pair of consumer cards, or a 64GB-plus Apple Silicon machine. That is demanding, but it is a different order of demand from the 405B model.
This does not make it a laptop model. On a single 24GB card you are pushed to a low quantisation and short context, so it is better understood as a recommendation for 48GB-and-up setups than a stretch goal for smaller machines.
The licensing
The Llama 3.3 Community License permits commercial use, with the same condition as recent Llama releases: organisations with more than 700 million monthly active users must request a separate licence from Meta, and attribution requirements apply. For the overwhelming majority of users this is not a practical constraint, but anyone deploying at consumer scale should read the terms rather than assume.