Nemotron 3 models
19 Nemotron 3 models from NVIDIA on Hugging Face, from 631M to 560.5B parameters, published by NVIDIA in F32 and BF16, with contexts from 256K to 1M tokens. The smallest official model, nemotron-3.5-asr-streaming-0.6b, needs about 2.9 GB of VRAM at its published precision; the cheapest live fit is Quadro M4000 at $0.046/hr.
Part of the Nemotron series
Pick a size
One row per official Nemotron 3 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| nemotron-3.5-asr-streaming-0.6b | 638M | 2.9 GB | 0.7 GB | 0.4 GB | Quadro M4000 | $0.046/hr | not published for this architecture |
| NVIDIA-Nemotron-3-Nano-4B-BF16 | 4.0B | 8.9 GB | 4.4 GB | 2.2 GB | RTX 5060 Ti | $0.145/hr | 0.50 GB |
| NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 31.6B | 70.6 GB | 35.3 GB | 17.6 GB | A100 | $1.31/hr | 0.19 GB |
| NVIDIA-Nemotron-3.5-Transcribe-30B-A4B | 33.7B | 75.3 GB | 37.7 GB | 18.8 GB | A100 | $1.31/hr | not published for this architecture |
| NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | 123.6B | 276 GB | 138 GB | 69.1 GB | RTX A6000 × 6 | $2.18/hr | 0.25 GB |
| NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | 560.5B | 1253 GB | 626 GB | 313 GB | AMD MI300X × 7 | $18.40/hr | not published for this architecture |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (10)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| nemotron-3.5-asr-streaming-0.6b | 638M | F32 | 2.9 GB | Quadro M4000 | $0.046/hr |
| NVIDIA-Nemotron-3-Nano-4B-BF16 | 4.0B | BF16 | 8.9 GB | RTX 5060 Ti | $0.145/hr |
| NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
| NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
| NVIDIA-Nemotron-3.5-Transcribe-30B-A4B | 33.7B | BF16 | 75.3 GB | A100 | $1.31/hr |
| NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16 | 123.6B | BF16 | 276 GB | RTX A6000 × 6 | $2.18/hr |
| NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | 123.6B | BF16 | 276 GB | RTX A6000 × 6 | $2.18/hr |
| NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | 560.5B | BF16 | 1253 GB | AMD MI300X × 7 | $18.40/hr |
Quantized versions (5)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark | 764M | BF16 | 1.7 GB | RTX 3070 | $0.088/hr |
| NVIDIA-Nemotron-3-Nano-4B-FP8 | 4.0B | F8_E4M3 | 4.4 GB | RTX 5060 Ti | $0.145/hr |
| NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 | 31.6B | F8_E4M3 | 35.3 GB | L40 | $0.759/hr |
| NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 123.6B | F8_E4M3 | 138 GB | AMD MI300X | $2.63/hr |
| NVIDIA-Nemotron-3-Ultra-550B-A55B-FP8-block | 560.5B | F8_E4M3 | 626 GB | RTX PRO 6000 × 7 | $9.63/hr |
Fine-tunes and community models (4)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| OpenThai2.0-OCR-Markdown | 631M | F32 | 2.8 GB | Quadro M4000 | $0.046/hr |
| nemotron3-8b | 8.5B | BF16 | 19.1 GB | RTX A5000 | $0.176/hr |
| Fastino-Nemotron-3.5-Lightning-Healthcare | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
| phonellm-alpha-1 | 31.6B | BF16 | 70.6 GB | A100 | $1.31/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune Nemotron 3
- How to run Nemotron 3: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Nemotron 3: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from NVIDIA
- Parakeet (8 models)
- Cosmos (7 models)
- Nemotron Nano (6 models)
- nvidia (2 models)
- Nemotron Labs (5 models)
- Nemotron Speech (1 models)
- Nemotron Parse (1 models)
- Nemotron-H (3 models)
- Nemotron Cascade (1 models)
- Canary (2 models)
- Nemotron VoiceChat (1 models)