NeuralStackMeta released Llama 4 on July 10 2026. The model supports a 10 million token context window and processes video natively at 60 frames per second. It achieved 92.4 percent on Video-MME and 89.1 percent on MMLU-Pro while running at 40 percent lower cost than GPT-5.1.
Training completed on July 2 using 128,000 H100 GPUs over 41 days. Meta published the full weights under a commercial license that allows fine-tuning and redistribution. Early adopters include Scale AI and Adept which integrated the model within 48 hours.
Background development began in January 2025 when Meta acquired two video-generation startups. The architecture combines a new mixture-of-experts router with temporal attention layers. Researchers trained on 28 trillion tokens including 4.2 million hours of video.
Previous Llama 3.1 model launched in July 2025 with only 128k context. Llama 4 triples parameter count to 1.8 trillion while improving efficiency through 16-way expert routing. The release includes a 70 billion parameter distilled variant for edge devices.
Why this matters
Enterprises gain open access to frontier-level video understanding without vendor lock-in. Cost reductions accelerate adoption in autonomous vehicles and content moderation. Regulators now face pressure to update open-source guidelines before the EU AI Act review in October 2026.
Meta plans quarterly updates through 2027 with planned support for 3D scene understanding. Developers can access the model immediately via Hugging Face and Meta's new inference API.