SmolLM — HuggingFace Compact Dense
HuggingFace's SmolLM2 tiny dense transformers. These deliver the engine's highest end-to-end throughput — SmolLM2-135M peaks at 662 tok/s on GGML-Vulkan.
Models
| Model | Params | 1BP Size | Backend(s) | Peak tok/s |
|---|---|---|---|---|
| SmolLM2-135M | 135M | 101 MiB | GGML-Vulkan / ZINC / CPU | 662 🏆 |
| SmolLM2-360M | 360M | 259 MiB | GGML-Vulkan / ZINC / CPU | 389 |
| SmolLM2-1.7B | 1.7B | 1007 MiB | GGML-Vulkan / ZINC / CPU | 167 |
Notes
- Measured 2026-08-01 on AMD Ryzen AI MAX+ 395 (Radeon 8060S, 32 GB UMA), all layers on Vulkan (
-ngl 999), Q4_K_M. - SmolLM2-135M also posts 3,646 tok/s prompt-prefill — the fastest prefill in the suite.
See also: benchmarks SSOT · full model support detail · all families