NeuralStackMeta released Llama 4 on July 18 2026 with a 10 million token context window and native video understanding capabilities. The model achieves 94.2 percent on MMLU benchmarks and delivers 3.8 times faster inference than Llama 3.1 405B on H100 clusters. Training completed on July 12 using 128,000 H100 GPUs over 42 days.
Weights are available under a new commercial license that permits fine-tuning for revenue-generating applications. Meta also open-sourced the full training dataset of 28 trillion tokens. Early access partners include Spotify and Accenture who began testing on July 19.
Background context includes Meta's prior Llama 3.1 release in July 2024 that topped open-source leaderboards for nine months. The company invested 18 billion dollars in AI infrastructure during 2025 to support larger models. Llama 4 training incorporated new safety filters developed after internal red-teaming exercises completed in June 2026.
Why this matters
The release accelerates the shift toward open-weight frontier models that enterprises can deploy without API costs. It pressures closed providers to justify premium pricing against freely available alternatives with comparable performance. Regulators are already examining the license terms for potential concentration risks.
Meta plans to ship Llama 4.1 with expanded multimodal support by October 2026. The move could reshape developer toolchains and reduce reliance on proprietary APIs across multiple industries.