1B- Context
- 8Ktokens
- Released
- 1y ago
- Size
- 1Bparameters
Developed by NVIDIA as part of its Nemotron retriever family, Llama Nemotron Embed 1B v2 is a text embedding model for retrieval, fine-tuned from Llama 3.2 1B. NVIDIA says it cuts data storage footprint by 35x through dynamic (Matryoshka) embedding sizing and long-document support. It is built for multilingual and cross-lingual question-answering retrieval over large text corpora.
How it compares
Additional source evidence
- Source model
- nvidia/llama-nemotron-embed-1b-v2
- Source family
- nemotron
- Source type
- embedding
- Last updated
- 2025-10-16
- Model context
- 8K
- Model output limit
- 2K
- Open weights
- Yes
- File attachments
- No
- Reasoning
- No
- Tool calling
- No
- Temperature control
- No
Downloadable weights
Complete source record (JSON)
{
"id": "nvidia/llama-nemotron-embed-1b-v2",
"name": "Llama Nemotron Embed 1B v2",
"description": "Multilingual text embedding model for retrieval and question answering pipelines",
"family": "nemotron",
"type": "embedding",
"last_updated": "2025-10-16",
"attachment": false,
"reasoning": false,
"tool_call": false,
"temperature": false,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 8192,
"output": 2048
},
"release_date": "2025-10-16",
"open_weights": true,
"license": "nvidia-open-model-license",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2"
}
]
}Frequently asked questions
Running locally
Llama Nemotron Embed 1B v2 needs an estimated 1.2 GB at Q4, 1.7 GB at Q8 or 2.8 GB at FP16. These are estimates that add about 15% over the weights plus 0.5 GB for a short context.
Yes, it fits on all four sizes at FP16 with an estimated 2.8 GB. So a 24 GB, 16 GB, 12 GB or 8 GB GPU is enough at that precision.
Yes. The Q4 estimate is 1.2 GB, and assuming macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits is 8 GB.
Privacy and data
Yes, you can self-host Llama Nemotron Embed 1B v2 for full control of your data because the model has open weights.
Capabilities and specs
The context window is 8K tokens (8,192) and the maximum output is 2K tokens (2,048).
No. The only input modality is text, so images, audio, video and PDFs are not supported.
No, tool calling (function calling) is not supported.
No, a reasoning or thinking mode is not supported.
Quality and benchmarks
The closest alternatives, matched on model kind and release date with no comparable price, are PaddleOCR-VL by Baidu, olmOCR 2 by Ai2 and ERNIE 4.5 VL 28B A3B Thinking by Baidu.
PaddleOCR-VLLlama Nemotron Embed 1B v2 vs PaddleOCR-VLolmOCR 2Llama Nemotron Embed 1B v2 vs olmOCR 2ERNIE 4.5 VL 28B A3B ThinkingLlama Nemotron Embed 1B v2 vs ERNIE 4.5 VL 28B A3B Thinking
Licensing and access
Llama Nemotron Embed 1B v2 has open weights. The license is nvidia-open-model-license.
The weights are available in the Hugging Face repository nvidia/llama-nemotron-embed-1b-v2.
Basics
Llama Nemotron Embed 1B v2 was made by NVIDIA and released on October 16, 2025.
Yes, Llama Nemotron Embed 1B v2 has open weights, so you can download and self-host it.
Generated from Vertical Tabs data, updated November 4, 2026.
Discussion
No comments yet
Be the first to share your thoughts.