Tech

Meta Just Gave Everyone a 405 Billion Parameter Model With No Strings Attached

June 21, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from ai.meta.com

What happened

Meta released Llama 3.1 405B, an open weights model trained on 15 trillion tokens that matches GPT-4 performance on MMLU, HumanEval, and GSM8K benchmarks. The weights are downloadable from Hugging Face and licensed for full commercial use with no usage caps. Researchers can now fine-tune the model on consumer GPUs or rent A100 clusters at roughly one tenth the cost of proprietary API calls.

Why it matters

You stop treating frontier models as black box services you rent by the token. Instead you treat them as infrastructure you can modify, host, and audit directly. This shifts your workflow from prompt iteration to actual model surgery and cost control.

Who's doing it

The Together AI research team fine-tuned Llama 3.1 405B on domain specific legal data and deployed it on their own inference stack. They report matching GPT-4 Turbo accuracy at 60 percent lower per token cost for their contract review product.

Try it

1. Visit huggingface.co/meta-llama/Meta-Llama-3.1-405B-Instruct and request access.
2. Download the weights using the Hugging Face CLI with your access token.
3. Run the model locally with vLLM on an 8xA100 node and test inference latency against your current API provider.

Read the original at ai.meta.com