Llama 3
Meta·
LLMsopen-weightWorkstation
Overview
Closed the quality gap with proprietary models significantly. The 70B variant rivaled GPT-3.5 Turbo on many tasks.
Capabilities and innovations
8B / 70B parameters8K context windowTiktoken-based 128K vocabularyStrong code and reasoning128K token vocabulary (4x Llama 2)15T token training corpusImproved post-training alignment
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 39.5% | unsourced |
| MMLU | 82% | unsourced |
Architecture and hardware
- Parameters
- 70B total (Dense)
- Estimated VRAM at Q4
- ~41 GB, Workstation class
- Quantization formats
- GGUF, GPTQ, AWQ
- Recommended runtime
- Ollama
- License
- Llama 3 Community License
Links
More from Meta
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.