Meta released Llama 3.1 405B in July 2024, marking the largest openly available language model trained from scratch. On standard academic benchmarks, the model demonstrates remarkable parity with GPT-4 and Claude 3 Opus: it achieves 85.2% on MMLU (expert knowledge across 57 subjects), compared to GPT-4's 86.4%, and scores 86.9% on GSM8K (grade-school math reasoning) versus GPT-4's 92%. On the MATH benchmark, Llama 405B reaches 53.3% accuracy—within striking distance of Opus's 60% and GPT-4's 52%. More importantly, the model handles long-context reasoning and code generation competitively, enabling tasks from legal document analysis to multi-file software debugging without reliance on external APIs.
The practical implications are substantial. A mid-sized financial services firm can now run the 405B model on eight NVIDIA H100 GPUs for approximately $150,000 in hardware—a one-time capital expenditure that replaces months of GPT-4 API spending (typically $10,000-50,000 monthly at scale). Using quantization techniques like 4-bit GPTQ, the model fits on consumer hardware with minimal performance loss, reducing the barrier to enterprise deployment. However, inference speed remains slower than cloud APIs: generating 100 tokens takes roughly 45 seconds on standard setups, making real-time chatbots less practical than batch applications like report generation or document analysis.
This release fundamentally shifts the open-source economics. Companies like Mistral, Together AI, and Hugging Face have built entire platforms around locally-deployable models, but Llama 405B's performance closes the quality gap that previously justified expensive API subscriptions. Anduril, a defense AI startup, deployed Llama 3.1 405B internally to avoid exfiltrating sensitive data through OpenAI's systems—a use case impossible before. The model's Apache 2.0 license removes legal friction for commercial use. This competitive pressure has already prompted OpenAI to lower API pricing and Anthropic to release faster inference options, suggesting the open-source ecosystem will continue commoditizing frontier AI capability.