Meta AI

Meta AI is the artificial intelligence research division of Meta Platforms (formerly Facebook). For local AI operators, Meta AI is most relevant as the developer of the Llama family of open-weight large language models (e.g., Llama 2, Llama 3, Llama 3.1). These models are released under permissive licenses and are widely used for local inference because they can be run on consumer hardware after quantization. Meta AI also produces other models like Segment Anything (image segmentation) and ImageBind (multimodal embeddings), but Llama models dominate local AI workflows.

Meta AI was formed in 2013 as the Facebook AI Research (FAIR) lab and later rebranded. Its mission is to advance AI research and open-source many of its models. The Llama series, starting with Llama 1 in 2023, set a new standard for open-weight LLMs. Llama 2 (2023) introduced commercial-friendly licensing, and Llama 3 (2024) improved performance and context length. Llama 3.1 (2024) added a 405B parameter model and extended context to 128K tokens. These models are popular in local AI because they are available in sizes from 8B to 405B, and quantization (e.g., Q4_K_M) makes them runnable on GPUs with 8–48 GB VRAM. Meta AI also contributes to tools like PyTorch and the LLaMA.cpp ecosystem, which are foundational for local inference.

An operator running Llama 3.1 8B via Ollama on an RTX 3060 (12 GB VRAM) can achieve ~30 tok/s at Q4_K_M quantization. The model is downloaded from Meta AI's official Hugging Face repository. If the operator tries the 70B variant, VRAM requirements jump to ~40 GB at Q4, requiring a 48 GB card or system-RAM offload.

In a typical workflow, an operator runs ollama pull llama3.1:8b to download the model from Meta AI's official distribution. The runtime loads the quantized weights into VRAM. If the operator wants to fine-tune the model, they download the full-precision weights from Hugging Face (e.g., meta-llama/Meta-Llama-3.1-8B) and use a tool like Unsloth or Axolotl.