Granite 4.0 · released 2025-10 · 3.19B parameters
Granite 4.0-H hybrid: 36 Mamba-2 layers (no KV cache) + 4 GQA attention layers (indices 5,15,25,35) with NoPE; every layer also has an MLP/MoE. Mamba state assumed BF16. Validated to 128K context. Dense MLP (8192).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.07 GB | 0.05 GB |
| 8,192 tokens | 0.10 GB | 0.07 GB |
| 32,768 tokens | 0.29 GB | 0.16 GB |
| 131,072 tokens | 1.04 GB | 0.54 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 5.94 GB |
| FP8 (W8A8) | 8 | 3.16 GB |
| INT8 (W8A8 / W8A16) | 8 | 3.16 GB |
| NVFP4 | 4.5 | 1.95 GB |
| INT4 AWQ / GPTQ | 4.25 | 1.86 GB |
| GGUF Q8_0 | 8.5 | 3.16 GB |
| GGUF Q6_K | 6.56 | 2.44 GB |
| GGUF Q5_K_M | 5.69 | 2.11 GB |
| GGUF Q4_K_M | 4.89 | 1.82 GB |
| GGUF Q3_K_M | 3.91 | 1.45 GB |
| GGUF Q2_K | 3.35 | 1.24 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× · 483 tok/s | 1× · 388 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 385 tok/s | 1× · 290 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 316 tok/s | 1× · 228 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 405 tok/s | 1× · 309 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 320 tok/s | 1× · 195 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 117 tok/s | 1× · 75.2 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 191 tok/s | 1× · 129 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 202 tok/s | 1× · 133 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 149 tok/s | 1× · 88.6 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 113 tok/s | 1× · 31.6 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 41.0 tok/s | 1× · 25.4 tok/s |