Meta has released Llama 4, its fourth-generation open-source language model, in three sizes: 8B, 70B, and 405B parameters. The 405B variant matches or exceeds GPT-4o across most standard benchmarks, marking a significant milestone for the open-source AI community.
The 70B model is particularly notable — it delivers 90% of the 405B model’s quality while running on a single consumer GPU (with quantization) or a small cluster of two A100s. This makes frontier-level AI accessible to startups, researchers, and hobbyists who can’t afford the $15+/M token pricing of proprietary APIs.
Benchmark Results
| Benchmark | Llama 4 405B | GPT-4o | Claude 3.5 Sonnet | |
|---|---|---|---|---|
| MMLU | 88.7% | 88.2% | 88.3% | |
| HumanEval | 92.1% | 90.2% | 92.0% | |
| MATH | 78.3% | 76.6% | 78.1% |
**Industry Impact**
The release has immediate implications for the AI industry. Companies that built products on GPT-4 or Claude can now switch to Llama 4 for a fraction of the cost by self-hosting. Several Y Combinator startups have already announced the switch.
Meta CEO Mark Zuckerberg framed the release as part of the company’s long-term strategy: “Open source AI is the foundation of the next generation of computing. We’re committed to ensuring these capabilities are available to everyone, not just the companies that can afford massive compute budgets.”