Meta Releases Llama 4: Open Source Multimodal Reasoning Shakes Up the Industry

Meta has officially released **Llama 4**, the next iteration of its open-source foundational model family. Available for immediate download on Hugging Face and for deployment across major cloud providers, Llama 4 introduces native multimodal vision support and embedded reasoning traces, delivering frontier performance to the open-source community for free.

The Llama 4 Family: 8B, 70B, and 405B

Following its previous strategy, Meta has structured Llama 4 into three distinct parameter sizes to serve different deployment environments:

• Llama 4-8B: Designed for edge devices, local laptops, and low-latency mobile apps. It features a compressed vision-language-action (VLA) block, enabling offline mobile spatial understanding.
• Llama 4-70B: The balanced workhorse model for corporate deployments and developer agents, offering high-tier coding and translation capabilities at manageable server costs.
• Llama 4-405B: The flagship model, matching or exceeding proprietary models like GPT-4o and Claude 3.5 Sonnet on complex reasoning, mathematics, and long-context analysis.

Native Vision and Reasoning Traces

Unlike previous Llama models that relied on external vision adapters, Llama 4 was pre-trained from the ground up on text, image, and video tokens. This native integration yields near-perfect visual OCR, allowing the model to extract and transcribe data from complex PDF charts and diagrams instantly.

Furthermore, Llama 4 features **Native Reasoning Traces**. During training, the model was taught to write its reasoning steps in a hidden scratchpad before delivering the final answer. This dramatically reduces logical errors in complex multi-step tasks like coding debugging and mathematical proofs.

Permissive Licensing for Businesses

Llama 4 is released under the **Meta Llama 4 License**, a permissive agreement that allows commercial modification and deployment for businesses with under 700 million monthly active users. This licensing strategy ensures that the vast majority of startups and enterprises can leverage the models without paying license fees, keeping pressure on proprietary API providers.