SmolLM · released 2025-07 · 3.08B parameters
NoPE on every 4th layer. 64K trained context (max_position_embeddings 65536); 128K with YaRN per card.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.28 GB | 0.14 GB |
| 8,192 tokens | 0.56 GB | 0.28 GB |
| 32,768 tokens | 2.25 GB | 1.13 GB |
| 131,072 tokens | 9.00 GB | 4.50 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 5.73 GB |
| FP8 (W8A8) | 8 | 3.11 GB |
| INT8 (W8A8 / W8A16) | 8 | 3.11 GB |
| NVFP4 | 4.5 | 1.96 GB |
| INT4 AWQ / GPTQ | 4.25 | 1.88 GB |
| GGUF Q8_0 | 8.5 | 3.04 GB |
| GGUF Q6_K | 6.56 | 2.35 GB |
| GGUF Q5_K_M | 5.69 | 2.04 GB |
| GGUF Q4_K_M | 4.89 | 1.75 GB |
| GGUF Q3_K_M | 3.91 | 1.40 GB |
| GGUF Q2_K | 3.35 | 1.20 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 1× · 486 tok/s | 1× · 239 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 383 tok/s | 1× · 162 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 312 tok/s | 1× · 121 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 403 tok/s | 1× · 176 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 303 tok/s | 1× · 85.8 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 112 tok/s | 1× · 35.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 186 tok/s | 1× · 62.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 199 tok/s | 1× · 62.0 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 144 tok/s | 1× · 37.5 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 108 tok/s | 1× · 25.1 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 39.3 tok/s | 1× · 11.4 tok/s |