LLaMA
Meta·
LLMsopen-weightWorkstation
Overview
First open-weight LLM from Meta, released to researchers. Proved that smaller open models could match much larger proprietary ones.
Capabilities and innovations
7B / 13B / 33B / 65B parametersCompetitive with GPT-3 at smaller scaleResearch-only licenseRMSNorm pre-normalizationRotary positional embeddings (RoPE)SwiGLU activation function
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| MMLU | 63.4% | unsourced |
Architecture and hardware
- Parameters
- 65B total (Dense)
- Estimated VRAM at Q4
- ~38 GB, Workstation class
- Quantization formats
- GGUF, GPTQ
- Recommended runtime
- llama.cpp
- License
- Research-only
Links
More from Meta
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.