---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.5-4B
tags:
- metis
- memory
- custom_code
- arxiv:2607.26760
---

# Metis-4B

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](https://huggingface.co/papers/2607.26760)
- [arXiv](https://arxiv.org/abs/2607.26760)
- [Code](https://github.com/MemTensor/Metis)

## 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.