Metis-4B is a Metis persistent-memory model built on Qwen/Qwen3.5-4B. The repository contains the complete merged model weights rather than a delta-only checkpoint.

Paper

This model is introduced in:

Metis: Memory Foundation Model

Hugging Face Paper
arXiv
Code

Load

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "IAAR-Shanghai/Metis-4B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
)

Metis uses custom Transformers code included in this repository. Use Transformers 5.4.0 or newer.