Meta releases Llama-3.1-405B weights with optimized inference

Full weights are available via Hugging Face and direct download. Benchmarks show 40 percent latency reduction using the new torch compile backend. Quantized versions for consumer GPUs are included in the repository.