Downloads · 30 days
0
olmoearth-nano-fp4
2imi9/olmoearth-nano-fp4
olmoearth-nano-fp4 is a feature extraction model from 2imi9. Use it when you need embeddings to search or compare text. It is set up for modelopt. The card lists the license as apache-2.0.
FP4 weight-quantized version of allenai/OlmoEarth-v1-Nano using nvidia-modelopt.
Access
Public
Updated Mar 25, 2026
Repo size
5.6 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pth5.6 MB · 100%
At a glance
- Task
- Feature Extraction
- Library
- modelopt
- License
- apache-2.0
- Access
- Public
- Created
- Mar 25, 2026
- Updated
- Mar 25, 2026
- SHA
- 6db05a3c
Base models
Model card
From the Hugging Face model README
OlmoEarth-v1-Nano (FP4 Quantized)
FP4 weight-quantized version of allenai/OlmoEarth-v1-Nano using nvidia-modelopt.
This is a quantized model, not a fine-tuned model.
Quantization Results
| Metric | Value |
|---|---|
| Quantization method | nvidia-modelopt FP4 PTQ |
| Cosine similarity vs FP32 (mean) | 0.978 |
| Cosine similarity vs FP32 (min) | 0.977 |
| Quantizer nodes | 111 (25 Linear + 12 Conv2d) |
| Save/load roundtrip | 1.000 (perfect) |
| Model size | 5.3 MB |
| GPU tested | RTX 5090 (Blackwell) |
Usage
import torch, copy
import modelopt.torch.opt as mto
from huggingface_hub import hf_hub_download
from olmoearth_pretrain.model_loader import ModelID, load_model_from_id
model = load_model_from_id(ModelID.OLMOEARTH_V1_NANO, load_weights=True).cuda().eval()
encoder = copy.deepcopy(model.encoder)
ckpt = hf_hub_download("2imi9/olmoearth-nano-fp4", "model.pth")
mto.restore(encoder, ckpt)
encoder.cuda().eval()
output = encoder(sample, patch_size=2)
embedding = output["project_aggregated"] # [B, 128]
Hardware
- FP4: Blackwell (RTX 5090, B100)
- CUDA 13.0+
- nvidia-modelopt >= 0.31
- Task
- Feature Extraction
- Library
- modelopt
- License
- apache-2.0
- Created
- Mar 25, 2026
- Updated
- Mar 25, 2026