Llama 3.1 405B is the flagship of Meta's Llama 3.1 family, with 405 billion parameters, and it launched on July 23, 2024 with refreshed 8B and 70B siblings. Meta described it as the "first frontier-level open source AI model", meaning its weights can be downloaded and run outside Meta's own products.

More than two years on, the 405B is no longer the newest Llama model: Meta has since released Llama 3.2 and 3.3. Its lasting importance lies in two things: the case that open weights can compete at the top, and its role in improving smaller models.

What Llama 3.1 405B added at launch

The pitch centers on parity. Across general knowledge, math, tool use and translation between languages, the 405B is presented as competitive with the best proprietary models (DevThrottle). Before this launch, Meta argued, open models typically lagged closed ones. Whether that gap has really closed deserves a fair test.

The smaller models share the 128K-token context window and support for eight languages (Meta). The license change is arguably the more consequential detail. Under the revised terms, a team can train a smaller system on the 405B's outputs, which opens the door to synthetic data and distillation (Hugging Face).

How was Llama 3.1 405B built and tested?

The training run reportedly covered upward of 15 trillion tokens, spread across more than 16,000 H100 chips (DevThrottle). The architecture stays conventional: a decoder-only transformer with modest tweaks, picked partly to keep training steady (Meta). Chat tuning repeated multiple passes of supervised fine-tuning, rejection sampling and direct preference optimization, and most fine-tuning examples were synthetic (Meta).

Llama 3.1 Model Architecture

Scores depend on how a test is run. A benchmark aggregator lists 88.6 on MMLU for the instruct model's chain-of-thought setting, while a Red Hat evaluation of the unquantized instruct model reports 87.41 under 5-shot prompting. Compare like with like before drawing conclusions.

Llama 3.1 performance

Can you run Llama 3.1 405B yourself?

Not on a typical computer. Meta compressed the model's numbers to 8-bit precision, which it says makes single-server deployment possible (Meta). On Hugging Face, the official 8-bit build is listed as running on eight H100 GPUs, and community 4-bit versions fit on eight A100 80GB cards. Meta says the weights are available from its own Llama site and from Hugging Face (Meta blog). For most individuals, the 8B model is the practical starting point. Our Meta Llama 3.1 page has more background.

Is Llama 3.1 405B still worth using?

For most users, a smaller model is the better choice. Hosting is shifting too. IONOS had scheduled the 405B for removal on September 15, 2026, and pointed users to Qwen3.5 397B A17B instead. That date has now passed, so check whether your provider still offers the model.

Its lasting value is as a teacher. Meta leaned on the 405B to lift the quality of its smaller siblings (Meta), and its results lent weight to the open-model side of the debate (DevThrottle). Benchmark parity alone does not settle whether the model fits your workload; the cost and hardware burden are real.

Key takeaways

  • Meta's parity claim rests on benchmark figures that shift with the test setup, so check results on your own tasks.
  • The license change, which lets outputs train other models, arguably matters as much as the scores.
  • The 405B is best treated as a specialist tool, such as a teacher model or research baseline, rather than a default choice.

Frequently asked questions

Is Llama 3.1 405B open source?

Its weights are downloadable, but the license requires any licensee whose products exceed 700 million monthly active users to ask Meta for permission. Because the license is custom rather than a standard open-source one, "open-weights" is the more precise label (memx).

Is Llama 3.1 405B still available?

Meta's Llama site still lists the 405B as an available model, though hosting providers may retire it.

Which Llama model should I try first?

Meta says Llama 3.3 70B approaches the 405B on some applications at much lower cost, which makes it the sensible first test for many projects (Meta).

Sources