Meta's Llama 3.1 405B, released in July 2024, is a landmark moment for open-source AI — the first freely available model that competes with frontier closed models.
**Benchmark Performance**
Llama 3.1 405B scores 88.6% on MMLU, comparable to GPT-4o. On math benchmarks (MATH) it scores 73.8%, and on coding (HumanEval) 89.0% — genuinely competitive with the best closed models.
**Context Window**
128K token context window across the entire 3.1 family, enabling long-document processing comparable to closed alternatives.
**Open Weights**
The model weights are freely downloadable under Meta's custom license (permissive for most commercial uses). This enables full local deployment with no API costs.
**Smaller Models**
The 8B and 70B variants offer excellent performance for their size. Llama 3.1 8B outperforms Llama 2 70B on most benchmarks, enabling capable AI on modest hardware.
**Hosting Options**
Available via Groq (very fast inference), Together AI, Fireworks AI, and others. Many providers offer it at $1-3 per million tokens.
**Verdict**
Llama 3.1 405B democratizes frontier AI. For organizations that cannot send data to OpenAI or Anthropic, it is the obvious choice.