Granite 4.1 · released 2026-04 · 8.38B parameters
Dense transformer. params from the base repo (tied embeddings); the instruct repo's safetensors (8.79B) additionally store a redundant lm_head copy.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.63 GB | 0.31 GB |
| 8,192 tokens | 1.25 GB | 0.63 GB |
| 32,768 tokens | 5.00 GB | 2.50 GB |
| 131,072 tokens | 20.0 GB | 10.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 15.6 GB |
| FP8 (W8A8) | 8 | 8.19 GB |
| INT8 (W8A8 / W8A16) | 8 | 8.19 GB |
| NVFP4 | 4.5 | 4.94 GB |
| INT4 AWQ / GPTQ | 4.25 | 4.71 GB |
| GGUF Q8_0 | 8.5 | 8.29 GB |
| GGUF Q6_K | 6.56 | 6.40 GB |
| GGUF Q5_K_M | 5.69 | 5.55 GB |
| GGUF Q4_K_M | 4.89 | 4.77 GB |
| GGUF Q3_K_M | 3.91 | 3.81 GB |
| GGUF Q2_K | 3.35 | 3.27 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× · 291 tok/s | 1× · 123 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 206 tok/s | 1× · 78.5 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 156 tok/s | 1× · 56.5 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 221 tok/s | 1× · 85.7 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 157 tok/s | 1× · 40.1 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 47.2 tok/s | 1× · 15.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 83.0 tok/s | 1× · 28.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 109 tok/s | 1× · 29.5 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 70.0 tok/s | 2× · 30.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 48.7 tok/s | 1× · 10.1 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 15.6 tok/s | 1× · 4.9 tok/s |