Meta Releases Llama 4 as Open Source

Meta has released Llama 4, its most powerful open-source language model to date, continuing the company's strategy of making frontier AI capabilities freely available. The model comes in two sizes — 405B and 70B parameters — and matches or exceeds the performance of proprietary models like GPT-4o across standard benchmarks.

Llama 4 405B achieves 90.1% on MMLU, 91.3% on HumanEval, and 94.2% on MATH, placing it firmly in frontier territory. The 70B variant, designed for deployment on more accessible hardware, scores within 5-8 percentage points of the larger model on most benchmarks, making it an attractive option for enterprises that need to run models on-premise.

Meta introduced several architectural innovations in Llama 4, including a new mixture-of-experts (MoE) design that activates only 52B parameters at inference time for the 405B model, dramatically reducing serving costs. The company also implemented a novel training approach called "constitutional self-play" that improves the model's ability to follow instructions while maintaining helpfulness.

The release includes fine-tuning recipes, quantized versions for consumer hardware, and integration guides for popular frameworks including vLLM, Hugging Face Transformers, and Ollama. Meta reports that Llama 4 70B can run efficiently on a single NVIDIA RTX 5090 GPU with 4-bit quantization, bringing near-frontier AI capabilities to individual developers.

The open-source AI community has responded enthusiastically, with over 50,000 downloads in the first 24 hours. Meta's VP of AI, Joelle Pineau, emphasized that open-source models are essential for AI safety research and for preventing any single company from controlling access to advanced AI capabilities.