Architecture: Mistral-based Genomic LLM
Quantization: AWQ (4-bit)
Primary Use: DNA sequence scoring, generation, and genomic feature analysis.

Benchmark Results (Local Evaluation)

Evaluation conducted on genomic sequences (max length 512) using TP=2.

Metric	FP16 (Original)	AWQ (4-bit)	Change
VRAM Footprint	~2x Model Size	12.5 GB	Optimized
Model Size	8112.6 MB	2143.4 MB	-73.6%
NLL Loss	5.7831	5.8353	+5.37% Drift
Perplexity (PPL)	324.7482	342.1725	+5.37% Drift
Generation Time	59.5s	66.4s	Optimized

Usage

Using vLLM

python

from vllm import LLM, SamplingParams

# Load the model
llm = LLM(model="ThomasYn/GenomeOcean-4B-AWQ", quantization="awq")

# Generate sequences
prompts = ["ATG", "GCA"]
sampling_params = SamplingParams(temperature=0.7, top_p=0.95, max_tokens=100)
outputs = llm.generate(prompts, sampling_params)

for output in outputs:
    print(f"Generated: {output.outputs[0].text}")

Using go-infer (GenomeOcean CLI)

bash

# Scoring sequences
python -m genomeocean.cli score --model_dir ThomasYn/GenomeOcean-4B-AWQ --sequence_file data.txt

# Generating sequences
python -m genomeocean.cli generate --model_dir ThomasYn/GenomeOcean-4B-AWQ --num 10 --max_seq_len 512

Model Repository Structure

This repository contains the necessary configuration files and model weights for AWQ inference.

model.safetensors: Quantized weights
config.json: Model configuration
modeling_mistral.py: Architecture implementation
tokenizer.json & tokenizer_config.json: Genomic tokenizer files

Citation

If you use this model in your research, please cite:

bibtex

@article{genomeocean2026,
  title={GenomeOcean: A Large-scale Foundation Model for Ocean Genomics},
  author={Thomas Yn, et al.},
  journal={bioRxiv},
  year={2026}
}