Meta releases 405 billion parameter Llama 3.1 with open weights
Meta published the full weights for Llama 3.1 405B. The model matches or exceeds GPT-4 on standard benchmarks and runs locally on consumer GPUs or inexpensive cloud instances. No per-token API fees apply.
Open weights shift the cost structure from usage fees to hardware and electricity. Users can now run large models without sending data to third-party servers. This changes deployment decisions from cloud subscription to infrastructure planning.
Hugging Face hosts the model weights and reports thousands of daily downloads. Independent developers have deployed quantized versions achieving 30 tokens per second on single RTX 4090 cards.
Step 1: Visit huggingface.co/meta-llama/Meta-Llama-3.1-405B and accept the license. Step 2: Install Ollama from ollama.com and run 'ollama run llama3.1:405b'. Step 3: Query the local endpoint and observe inference speed without API costs.