Llama — Meta Dense Transformers
Meta's dense transformers with broad backend coverage — GGUF runs on every GPU backend and CPU, with NPU support for the 1B/3B/8B tier.
Models
| Model | Params | 1BP Size | Backend(s) | Perf |
|---|---|---|---|---|
| Llama-3.2-1B | 1B | 581 MB | GGML-Vulkan / ZINC / NPU | — |
| Llama-3.2-3B | 3B | 1.7 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Llama-3.1-8B | 8B | 4.1 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| TinyLlama-1.1B | 1.1B | 328 MB | GGML-Vulkan / ZINC / NPU | — |
Notes
- NPU: Llama3.2 1B, Llama3.2 3B, Llama3.1 8B — build with
./build_xclbins.sh llama. - GPU HIP / Vulkan / CPU: GGUF validated across all.
- Llama is also the base architecture for DeepSeek-R1-Distill-Llama-8B (44 tok/s GGML-Vulkan) — see DeepSeek page.
See also: full model support detail · benchmarks SSOT · all families