Nemotron 3 · released 2026-03 · 3.97B parameters
Nemotron-H hybrid: each block is a single Mamba-2 mixer, an MLP/MoE, or a GQA attention layer; only attention blocks keep a KV cache. Compressed from Nano 9B v2 (Nemotron Elastic): 21 Mamba-2 + 17 MLP + 4 attention blocks. Mamba state assumed BF16.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.10 GB | 0.07 GB |
| 8,192 tokens | 0.16 GB | 0.10 GB |
| 32,768 tokens | 0.54 GB | 0.29 GB |
| 131,072 tokens | 2.04 GB | 1.04 GB |
| 262,144 tokens | 4.04 GB | 2.04 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 7.40 GB |
| FP8 (W8A8) | 8 | 5.23 GB |
| INT8 (W8A8 / W8A16) | 8 | 5.23 GB |
| NVFP4 | 4.5 | 4.28 GB |
| INT4 AWQ / GPTQ | 4.25 | 4.21 GB |
| GGUF Q8_0 | 8.5 | 3.93 GB |
| GGUF Q6_K | 6.56 | 3.03 GB |
| GGUF Q5_K_M | 5.69 | 2.63 GB |
| GGUF Q4_K_M | 4.89 | 2.26 GB |
| GGUF Q3_K_M | 3.91 | 1.81 GB |
| GGUF Q2_K | 3.35 | 1.55 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× · 384 tok/s | 1× · 296 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 288 tok/s | 1× · 210 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 227 tok/s | 1× · 160 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 306 tok/s | 1× · 226 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 184 tok/s | 1× · 118 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 74.9 tok/s | 1× · 48.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 128 tok/s | 1× · 85.5 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 179 tok/s | 1× · 105 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 128 tok/s | 1× · 67.6 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 95.0 tok/s | 1× · 24.9 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 25.3 tok/s | 1× · 16.1 tok/s |