Nucleotide Transformer 2.5B NF4

This model is an NF4 (Normal Float 4-bit) quantized version of the base model InstaDeepAI/nucleotide-transformer-2.5b-multi-species. The checkpoint was quantized using the BitsAndBytes library with double quantization enabled and BF16 computation.

The goal of this release is to significantly reduce memory and storage requirements while preserving the representation quality of the original model, making large-scale genomic foundation models more accessible on consumer and research hardware.

Base Model

This checkpoint is derived from:

• InstaDeepAI/nucleotide-transformer-2.5b-multi-species

Please refer to the original model card for details regarding training data, architecture, intended use cases, limitations, and downstream biological applications.

Quantization Details

• Quantization Method: BitsAndBytes NF4
• Precision: 4-bit
• Quantization Type: NF4 (Normal Float 4)
• Double Quantization: Enabled
• Compute Dtype: BF16

Benchmark Summary

The quantized model was evaluated against the original BF16 model using hidden-state similarity and embedding preservation metrics.

These results indicate that the quantized model closely preserves the representations produced by the original checkpoint while substantially reducing memory requirements.

Intended Use

This model is suitable for:

• Genomic sequence embeddings
• DNA representation learning
• Biological sequence analysis
• Taxonomic and biodiversity research
• Downstream bioinformatics tasks
• Resource-constrained inference environments

Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer

model_id = "divyeshkamalanaban/nucleotide-transformer-2.5b-nf4"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForMaskedLM.from_pretrained(
 model_id,
 device_map="auto",
)

sequence = "ACGT" * 250

inputs = tokenizer(
 sequence,
 return_tensors="pt",
)

outputs = model(**inputs)

Acknowledgements

This work builds upon the Nucleotide Transformer project developed by InstaDeep and collaborators. All credit for the original model architecture, training pipeline, datasets, and research contributions belongs to the original authors.

This repository provides a community-maintained NF4 quantized checkpoint intended to improve deployment efficiency and accessibility.

Base model

InstaDeepAI/nucleotide-transformer-2.5b-multi-species

(1)