NeuralStackMeta stated on July 27 2026 that Llama 4 will be released with fully open weights on August 12. The 405 billion parameter model supports text image and audio inputs and can run inference on a single H100 GPU at 32 tokens per second. The license permits commercial use and fine-tuning without restrictions.
Meta trained Llama 4 on 15 trillion tokens including 2.4 trillion image-text pairs. Internal evaluations show the model matches or exceeds GPT-5 on 11 of 14 standard benchmarks while using 38 percent less compute at inference.
The company will also release a distilled 70 billion parameter version optimized for edge devices. Meta expects over 50 million downloads in the first month based on Llama 3 adoption patterns.
Meta has released three generations of Llama models since 2023 and has positioned open-source releases as a counterweight to closed API providers. The company maintains that open weights accelerate safety research through community scrutiny.
Hardware partners including Nvidia and AMD have prepared day-one support for the new model architecture.
Why this matters
Llama 4's open release lowers barriers for startups and researchers who cannot afford proprietary API costs. Widespread fine-tuning could produce specialized models for non-English languages and niche domains within weeks.
The move pressures closed-model companies to justify premium pricing and may influence upcoming AI regulation debates around open-source access. Developers gain a high-performance baseline they can audit and modify directly.
Meta's strategy continues to shape the competitive landscape by commoditizing frontier model capabilities.