Meta drops Llama 3.1 405B to challenge GPT-4o

Meta's Llama 3.1 is an open-weight model family featuring three sizes: 8B, 70B, and a massive 405B parameter flagship. Every size comes in base and instruction-tuned variants, all upgraded with a 128K token context window and native support for 8 languages.

The biggest deal here is the updated license. Meta officially allows developers to use outputs from Llama 3.1—including the 405B powerhouse—to train, improve, and distill other AI models, lifting a major historical restriction on open development.

What it can do

The 405B flagship stands as the largest openly available dense foundation model to date. Across more than 150 benchmark datasets and blind human evaluations, Meta shows that 405B goes toe-to-toe with proprietary leaders like GPT-4, GPT-4o, and Claude 3.5 Sonnet in coding, math, and general reasoning.

The instruct models have also gained agentic tool-use capabilities. They come with built-in integrations for web search and Wolfram Alpha, and can execute custom JSON function calls.

Training the 405B model required over 16,000 Nvidia H100 GPUs churning through more than 15 trillion tokens, contributing to a total of 39.3 million GPU hours across the Llama 3.1 collection. Alongside the core models, Meta also released Llama Guard 3 for safety moderation and Prompt Guard, a lightweight 279M classifier built specifically to intercept prompt injections.

Hardware requirements and where to try it

Hardware demands depend heavily on which model you run. The lightweight 8B model fits easily on consumer-grade GPUs, and you can even fine-tune it on a single GPU using Hugging Face's TRL library.

The mid-sized 70B targets enterprise setups, while running the 405B model locally remains a heavy engineering task. In standard 16-bit precision, it demands multi-node GPU clusters, but Meta introduced an FP8 quantized version that allows the 405B model to run inference within a single 8-GPU server node.

You can download weights directly from Hugging Face or llama.meta.com. If you want to try 405B without managing infrastructure, it is accessible on Hugging Chat, with API endpoints hosted across partners like Groq, AWS, and Google Cloud.

Why it matters

For most developers, you do not actually need to host the 405B model yourself. The real game-changer is using 405B as a teacher: generating synthetic data, evaluating other outputs, and distilling top-tier performance into lean 8B models without legal gray areas.

We think this release fundamentally changes the economics of foundation models. By offering GPT-4o parity under an open license that encourages distillation, Meta has dealt a severe blow to closed-source API margins and handed builders an unprecedented level of independence.