Nemotron 3 · released 2026-08 · 31.6B parameters (3.00B active per token)
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. MoE blocks: 128 experts top-6 + 1 shared. Mamba SSM cache FP32 per config. Same block layout as Nemotron 3 Nano 30B-A3B (+1 MTP layer in config); card: 30B / 3B active, up to 1M context.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.07 GB | 0.06 GB |
| 8,192 tokens | 0.09 GB | 0.07 GB |
| 32,768 tokens | 0.23 GB | 0.14 GB |
| 131,072 tokens | 0.80 GB | 0.42 GB |
| 1,048,576 tokens | 6.05 GB | 3.05 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 58.8 GB |
| FP8 (W8A8) | 8 | 30.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 30.7 GB |
| NVFP4 | 4.5 | 18.4 GB |
| INT4 AWQ / GPTQ | 4.25 | 17.6 GB |
| GGUF Q8_0 | 8.5 | 31.2 GB |
| GGUF Q6_K | 6.56 | 24.1 GB |
| GGUF Q5_K_M | 5.69 | 20.9 GB |
| GGUF Q4_K_M | 4.89 | 18.0 GB |
| GGUF Q3_K_M | 3.91 | 14.4 GB |
| GGUF Q2_K | 3.35 | 12.3 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× · 526 tok/s | 1× · 127 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 451 tok/s | 1× · 82.2 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 391 tok/s | 1× · 59.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 467 tok/s | 1× · 89.6 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 397 tok/s | 1× · 59.6 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 172 tok/s | 1× · 16.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 263 tok/s | 1× · 29.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 224 tok/s | 1× · 44.0 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 180 tok/s | 2× · 41.8 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 145 tok/s | 1× · 17.6 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 65.1 tok/s | 1× · 5.3 tok/s |