// NVIDIA Nemotron · Omni / audio LLM
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16: VRAM & local deployment
33B total · 3B active (MoE) parameters. About 22.6 GB at 4-bit and 73.2 GB at BF16 with one 8K-token request.
released 2026-04-20licence: Commercial with conditionsmemory figures are estimates
01Specs
| Publisher | NVIDIA · nvidia |
|---|---|
| Family / series | NVIDIA Nemotron · Nemotron-3-Nano-Omni-Reasoning |
| Type | Omni / audio LLM (any-to-any) |
| Parameters | 33B total · 3B active (MoE) Source: counted from official safetensors weights |
| Context length | 256K tokens (262,144)max_position_embeddings in config.json |
| Architecture | 52 layers (6 full-attention) · GQA+HYBRID · MoE |
| Released weights | BF16 · 66.0 GBdownload |
| Licence | nvidia-open-model-agreement · Commercial with conditionsCommercial use allowed under the publisher’s own licence terms (user caps, acceptable-use policy, attribution or region limits). Read it before shipping. |
| Released | 2026-04-20 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Total @ 128K | Smallest GPU | Mac |
|---|---|---|---|---|---|---|
| BF16 / FP16 (original) | 66.0 GB | 73.2 GB | 73.3 GB | 73.9 GB | RTX PRO 6000 96GB | Mac 128GB |
| FP8 / INT8 / Q8_0 | 35.1 GB | 39.1 GB | 39.3 GB | 39.9 GB | 2 × RTX 5090 (64GB) | Mac 64GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 20.0 GB | 22.6 GB | 22.7 GB | 23.3 GB | RTX 5090 32GB | Mac 64GB |
KV cache per request: 8K 0.1 GB · 32K 0.2 GB · 128K 0.8 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- vLLM — mentioned in the official model card
- SGLang — mentioned in the official model card
- llama.cpp — mentioned in the official model card
- Ollama — mentioned in the official model card
- Transformers — mentioned in the official model card
- TensorRT-LLM — mentioned in the official model card
04Official repositories & quantised builds
3 official repo(s) from nvidia.
05Other NVIDIA Nemotron models
All 62 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 | 3.98B | 3.6 GB | 2026-10-16 |
| Nemotron-3.5-Rerank-8B-BF16 | 8.38B | 6.1 GB | 2026-10-15 |
| NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | 124.5B / 12B | 83.5 GB | 2026-10-12 |
| NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 | 74.9B / 9B | 50.4 GB | 2026-10-09 |
| Nemotron-3.5-Embed-1B-BF16 | 1.55B | 1.5 GB | 2026-10-07 |
| Nemotron-3.5-Embed-8B-BF16 | 8.38B | 6.1 GB | 2026-10-07 |
| Nemotron-3.5-Parse | 631M | 1.1 GB | 2026-09-03 |
| Nemotron-3-Labs-Ultra-Math-RL | 550B / 55B | 367 GB | 2026-09-02 |
06Sources
Data checked 2026-08-24. Parameters, context and licence are copied from the official repository; memory figures are our estimates.