Nemotron 3.5 Parse
Parameters
630M
Context Length
16K
Modality
Text
Architecture
Dense
KV Cache Rate
20 KB/tok
License
openmdw-1.1
Release Date
3 Sept 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Self-hosted only
Hardware & System Requirements
Precision
Context
Workload Memory: 2.95 GB
INT4 · 8K context
Base Weights: 0.29 GB
KV Cache: 0.16 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
1x NVIDIA L4 24GB
2.9 / 21.6 GB usable
1x RTX 3060 12GB
2.9 / 10.8 GB usable
Apple M4 24GB
2.9 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Runtime requirements
Checkpoint format
BF16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 5.6.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native BF16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native BF16 | FlashAttention-2 |
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for Nemotron 3.5 Parse available.
Rankings
Overall Rank
-
Coding Rank
-
About Nemotron 3.5 Parse
Nemotron 3.5 Parse is a specialized foundation parser developed by NVIDIA for complex document understanding and structured information extraction. It leverages advanced prefix causal attention to process dense visual and textual documents.
Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
16
Key-Value Heads
8
Attention Head Dimension
64
Position Embedding
Absolute Position Embedding
Normalization
Layer Normalization
Activation Function
GELU
Dimensions
Base Precision
BF16
Hidden Dimension Size
1,024
Number of Layers
10
FFN Intermediate Size (Dense)
4,096
Multi-Token Prediction Heads
1
Tokenizer
Vocabulary Size
72,512
Resources
About Nemotron 3.5
NVIDIA's Nemotron 3.5 family of open models delivers leading efficiency and accuracy for agentic AI applications. Built on an advanced hybrid Mamba-2 SSM and sparse Mixture-of-Experts architecture with native Multi-Token Prediction (MTP) and speculative decoding harnesses, Nemotron 3.5 models are purpose-built for the high-volume execution layer of always-on agents.