Meta drops Llama 3 with 8B and 70B open weights

Llama 3 is Meta’s next-generation family of open-access large language models, launching in two core sizes: 8B and 70B parameters. Both sizes are available as raw base checkpoints and dialogue-optimized Instruct models, accompanied by Llama Guard 2, a safety classifier fine-tuned on the 8B architecture.

All models are now live on the Hugging Face Hub with full ecosystem support, ranging from Hugging Chat web demos to single-GPU fine-tuning scripts.

Specs, hardware, and performance

Meta built Llama 3 on an expanded scale. Training ran across two 24,000-GPU clusters using over 15 trillion tokens of public online data, roughly an eightfold increase over Llama 2. For the Instruct checkpoints, Meta layered on more than 10 million human-annotated examples using SFT, rejection sampling, PPO, and DPO.

Under the hood, Llama 3 upgrades its tokenizer with a massive 128,256 vocabulary, up from 32,000 in Llama 2. This expanded vocabulary compresses text more efficiently and boosts multilingual performance, though the larger embedding tables account for the small model growing from 7B to 8B parameters. The 8B variant also gains Grouped-Query Attention (GQA) for improved inference efficiency. Context window length currently sits at 8k tokens across all variants.

On the Open LLM Leaderboard, Llama 3 posts strong numbers. The 8B version scored 13.41, handily beating Llama 2 7B at 8.72. The 70B version reached 26.37, pulling past Llama 2 70B at 18.25 and edging out Mixtral 8x22B at 25.49.

Resource requirements remain approachable. Llama 3 8B Instruct in bfloat16 consumes around 16 GB of VRAM, running comfortably on desktop GPUs like the RTX 3090 or RTX 4090. Quantized down to 4-bit, memory requirements fall to roughly 7 GB, putting it well within reach of budget GPUs and free Google Colab tiers.

Deployment and licensing

Developers can run Llama 3 locally using Transformers 4.40 or higher, with torch.compile and CUDA graphs delivering up to a 4x inference speedup. For production workloads, one-click deployments are live on Hugging Face Inference Endpoints, AWS SageMaker, and Google Cloud via Vertex AI or GKE. A live web demo of the 70B Instruct model is also available on Hugging Chat.

The Llama 3 license remains commercially permissive, allowing redistribution, fine-tuning, and derivative works. However, Meta added stricter attribution rules: derived models must include Llama 3 at the start of their name, and downstream apps or services must explicitly state they were Built with Meta Llama 3.

Why it matters

For builders and indie developers, Llama 3 8B is the new default starter pack. You get frontier-grade reasoning that fits in 7 GB of VRAM, single-GPU QLoRA adaptability, and performance that surpasses older 13B class models.

For the AI landscape, Meta just proved that pouring 15 trillion tokens of high-quality data into dense architectures can effectively bridge the gap between open weights and proprietary cloud APIs. The 8k context window feels tight in a world drifting toward millions of tokens, but as an open, customizable foundation for real-world products, Llama 3 resets the open-source bar.