Nemotron 3 · released 2026-06 · 561B parameters (55.0B 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. 48 Mamba-2 + 48 LatentMoE (512 experts top-22, 2048-dim latent) + 12 attention blocks; +1 MTP. Block counts derived from the exact parameter total and the 49 FP32 router-bias tensors (summarised config listings were inconsistent). Card: 550B / 55B active, up to 1M context. NVFP4 checkpoint published.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.43 GB | 0.40 GB |
| 8,192 tokens | 0.47 GB | 0.43 GB |
| 32,768 tokens | 0.75 GB | 0.57 GB |
| 131,072 tokens | 1.88 GB | 1.13 GB |
| 1,048,576 tokens | 12.4 GB | 6.38 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1,044.1 GB |
| FP8 (W8A8) | 8 | 526.0 GB |
| INT8 (W8A8 / W8A16) | 8 | 526.0 GB |
| NVFP4 | 4.5 | 299.4 GB |
| INT4 AWQ / GPTQ | 4.25 | 283.2 GB |
| GGUF Q8_0 | 8.5 | 554.7 GB |
| GGUF Q6_K | 6.56 | 428.1 GB |
| GGUF Q5_K_M | 5.69 | 371.3 GB |
| GGUF Q4_K_M | 4.89 | 319.1 GB |
| GGUF Q3_K_M | 3.91 | 255.1 GB |
| GGUF Q2_K | 3.35 | 218.6 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) | 4× · 95.7 tok/s | 4× · 29.4 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 8× · 53.2 tok/s | 8× · 25.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 16× · 12.3 tok/s | 16× · 9.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 87.7 tok/s | 4× · 21.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 8× · 52.0 tok/s | 8× · 21.9 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 11.8 tok/s | 16× · 6.7 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 23.1 tok/s | 8× · 9.3 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 11.9 tok/s | 16× · 9.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | > 16 | > 16 |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |