Meta releases Llama 3.1 405B under open license
Meta published the weights for its 405 billion parameter Llama 3.1 model with a permissive open license. Users can download, run, and fine tune the model on local hardware without incurring API charges. The release includes training code, evaluation benchmarks, and safety reports.
This removes the paywall that previously limited frontier scale models to large corporations. Teams can now experiment with model scale directly instead of relying on rate limited APIs. The change shifts planning from token budgets to hardware and electricity budgets.
Hugging Face hosts the model weights and provides one click deployment templates. Their public leaderboard shows over 120,000 downloads and multiple fine tuned variants already outperforming closed models on academic benchmarks.
Step 1: Visit huggingface.co/meta-llama/Meta-Llama-3.1-405B and accept the license. Step 2: Use the transformers library command pip install transformers and load the model with from_pretrained. Step 3: Run inference on a single A100 GPU and record tokens per second and memory usage.