Parameters
4.0B
Architecture
Mamba2-Transformer Hybrid (Nemotron-Hybrid)
Released
16.03.2026
License
NVIDIA Nemotron Open Model License
Input Modalities
Output Modalities
Context (native)
262,144 tokens
Context (extended)
262,144 tokens
About
NVIDIA-Nemotron-3-Nano-4B is a small language model (SLM) trained from scratch by NVIDIA and designed as a unified model for both reasoning and non-reasoning tasks: it answers queries by first generating a reasoning trace, then concluding with a final response, with the reasoning behavior controllable via a system prompt. It was compressed from NVIDIA-Nemotron-Nano-9B-v2 using the Nemotron Elastic framework (arXiv 2511.16664).
Its Mamba2-Transformer hybrid (Nemotron-Hybrid) architecture consists primarily of Mamba2 SSM and MLP layers combined with just four Attention layers, giving 3.97B parameters with a 262K-token context window and high inference efficiency on NVIDIA hardware (A10G, A100, H100, GeForce RTX; NeMo 25.07 runtime). Input and output are text, English and coding languages.
Pre-training used more than 10 trillion tokens (data cutoff September 2024). The post-training corpus spans English and multilingual text across code, legal, math, science and finance domains, and includes synthetic reasoning traces distilled from DeepSeek R1/R1-0528, Qwen3-235B-A22B, Nemotron 4 340B and Qwen2.5 models - the model notes it was "Improved using Qwen".
It targets edge-ready Agentic AI on Jetson Thor, GeForce RTX and DGX Spark: AI gaming NPCs (teammates/companions), local voice assistants, and IoT automation. Model dates Dec 2025 - Jan 2026; released on Hugging Face March 16, 2026 under the NVIDIA Nemotron Open Model License; ready for commercial use.
Training Data 10+ trillion tokens pre-training, compressed from NVIDIA-Nemotron-Nano-9B-v2
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
BFCL-V4
general_agent
|
31.99%
|
— |
#23/27
better than 15%
|
|
MMLU-Pro
knowledge
|
18.39%
|
— |
#42/48
better than 13%
|
|
SWE-bench Verified
coding_agent
|
2.98%
|
— |
#58/60
better than 3%
|
|
Humanity's Last Exam
stem_reasoning
|
5.30%
|
— |
#68/71
better than 4%
|
|
GPQA Diamond
stem_reasoning
|
18.70%
|
— |
#73/80
better than 8%
|
|
Terminal Bench 2.1
coding_agent
|
3.77%
|
— |
#40/44
better than 9%
|
|
SuperGPQA
knowledge
|
26.13%
|
— |
#25/28
better than 11%
|
|
BrowseComp-zh
general_agent
|
3.30
|
— |
#13/13
better than 0%
|
|
BrowseComp
general_agent
|
1.80%
|
— |
#43/45
better than 5%
|
|
NoLiMa
long_context
|
0.89%
|
— |
#7/9
better than 25%
|
|
Gaia2
general_agent
|
26.50
|
— |
#12/12
better than 0%
|
|
GDPVal-AA v2
general_agent
|
0.00
|
— |
#33/38
better than 0%
|
|
Claw-Eval Avg
coding_agent
|
38.90%
|
— |
#35/40
better than 13%
|
|
LongBench v2
long_context
|
16.06%
|
— |
#19/22
better than 14%
|
|
WildClawBench
coding_agent
|
6.47%
|
— |
#18/20
better than 5%
|
|
TAU3-Bench
general_agent
|
1.20
|
— |
#21/21
better than 0%
|
|
QwenClawBench
coding_agent
|
25.15%
|
— |
#14/17
better than 19%
|
|
TAU2-Bench
general_agent
|
14.81%
|
— |
#18/21
better than 15%
|
|
LiveCodeBench v6
stem_reasoning
|
41.61%
|
— |
#43/49
better than 12%
|
|
LCB-Pro 25Q2 (Easy)
stem_reasoning
|
71.58%
|
— |
#4/9
better than 62%
|
|
LCB-Pro 25Q2 (Medium)
stem_reasoning
|
30.29%
|
— |
#3/9
better than 62%
|
|
OJBench
stem_reasoning
|
25.55%
|
— |
#17/22
better than 24%
|
|
SciCode
|
19.81%
|
— |
#24/27
better than 12%
|
|
AIME 26
stem_reasoning
|
52.68%
|
— |
#39/45
better than 14%
|
|
HMMT Feb 26
stem_reasoning
|
39.90%
|
— |
#38/44
better than 14%
|
|
MATH-500
math
|
45.59%
|
— |
#7/11
better than 40%
|
|
IFBench
instruction_following
|
57.89%
|
— |
#37/42
better than 12%
|
|
MMLU-Redux
knowledge
|
33.61%
|
— |
#28/30
better than 7%
|
|
LongBenchPro
long_context
|
34.99%
|
— |
#7/10
better than 33%
|
|
AIME 2025
stem_reasoning
|
37.93%
|
— |
#19/24
better than 22%
|
|
Multi-IF
instruction_following
|
52.46%
|
— |
#10/13
better than 25%
|
|
IFEval
instruction_following
|
85.88%
|
— |
#19/26
better than 28%
|
|
SWE-bench Pro
coding_agent
|
0.11%
|
— |
#76/77
better than 1%
|
|
AA-LCR
long_context
|
20.60%
|
— |
#40/42
better than 5%
|
What is a good BFCL-V4 score? What is a good MMLU-Pro score? What is a good SWE-bench Verified score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 7.9 GB | 0.6 GB | 9.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 4.0 GB | 0.3 GB | 4.5 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 2.0 GB | 0.2 GB | 2.2 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 2.0 GB | 0.2 GB | 2.2 GB | Which GPUs can run it? Calculate on your GPU |
Consumer GPUs that fit this model with a single card (INT4):
Architecture
- Attention
- Hybrid Attention (40:8)
- Layers
- 42
- Hidden size
- 3136
- Context
- 262K tokens
- Parameters
- 3970M
Source: Hugging Face config.json · NemotronHForCausalLM · model repo
Nemotron-Hybrid
Nemotron Elastic framework (arXiv 2511.16664)
A10G, A100, H100-80GB, GeForce RTX, Jetson Thor, DGX Spark
text
text
NVIDIA-Nemotron-Nano-9B-v2
NeMo 25.07
Training Pipeline
-
1
pretraining
Pre-training (>10T tokens)
Trained from scratch on more than 10 trillion tokens (data cutoff September 2024).
-
2
other
Elastic compression from Nemotron-Nano-9B-v2
Compressed from the 9B parent via the Nemotron Elastic framework (arXiv 2511.16664).
-
3
sft
Post-training SFT with synthetic reasoning traces
English + multilingual corpus across code, legal, math, science, finance; synthetic reasoning traces distilled from DeepSeek R1/R1-0528, Qwen3-235B-A22B, Nemotron 4 340B, Qwen2.5-32B/14B/72B; improved using Qwen.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| AI2's Reasoning Challenge | finetune | — | — | |
| APIGen Function-Calling | finetune | — | — | |
| Bespoke-Stratos-17k | finetune | — | — | |
| Big-Math-RL-Verified | finetune | — | — | |
| Cosmos QA | finetune | — | — | |
| Essential-Web | pretraining | — | — | |
| FineMath | pretraining | — | — | |
| FineWeb-2 | pretraining | — | — | |
| HelpSteer3 | finetune | — | — | |
| LMSYS-Chat-1M | finetune | — | — | |
| MCTest | finetune | — | — | |
| MMLU Auxiliary Train | finetune | — | — | |
| MedMCQA | finetune | — | — | |
| MegaMath | pretraining | — | — | |
| MetaMathQA | finetune | — | — | |
| Moral Stories | finetune | — | — | |
| MultiverseMathHard | finetune | — | — | |
| NuminaMath CoT | finetune | — | — | |
| OpenCodeReasoning-2 | finetune | — | — | |
| OpenWebMath | pretraining | — | — | |
| SWE-Gym | finetune | — | — | |
| Skywork-OR1-RL-Data | finetune | — | — | |
| The Stack | pretraining | — | — | |
| WikiTableQuestions | finetune | — | — | |
| WildChat-1M | finetune | — | — | |
| arithmetic | finetune | — | — | |
| finepdfs | pretraining | — | — | |
| glaive-function-calling-v2 | finetune | — | — | |
| mC4 | pretraining | — | — | |
| opc-sft-stage2 | finetune | — | — | |
| peS2o | pretraining | — | — | |
| simple-arithmetic-problems | finetune | — | — | |
| social-chemestry-101 | finetune | — | — | |
| tigerbot-kaggle-leetcodesolutions-en-2k | finetune | — | — | |
| Private third-party datasets (Global Regulation, Workbench) | training | — | — | |
| Online dataset sources (Common Crawl Foundation) | pretraining | — | — | |
| NVIDIA-sourced synthetic datasets | training | — | — |
Linked Resources
NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model
https://research.nvidia.com/labs/adlr/files/NVIDIA-Nemotron-Nano-2-Technical-Report.pdf
Elastic model compression framework (Nemotron Elastic)
https://arxiv.org/abs/2511.16664
Nemotron-H tech report
https://arxiv.org/abs/2504.03624
NVIDIA/NeMo-Skills (evaluation harness)
https://github.com/NVIDIA/NeMo-Skills
Trend Analysis
24h Change
+1.4%
7d Change
+2.1%
Current
143
followers
+0.1%
downloads
+0.0%
downloads
+0.4%
downloads_all_time
+0.1%
No comments yet
Be the first to share your thoughts!