README.md

---
license: apache-2.0
---

```python
import os
import torch
from transformers import (
    AutoConfig,
    AutoModelForImageTextToText,
    AutoProcessor,
    AutoTokenizer,
)

model_id = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
torch.manual_seed(42)
config = AutoConfig.from_pretrained(model_id)

config.tie_word_embeddings = False
config.text_config.tie_word_embeddings = False

config.text_config.num_hidden_layers = 2
config.text_config.hidden_size = 64
config.text_config.intermediate_size = 128
config.text_config.num_attention_heads = 4
config.text_config.num_key_value_heads = 2
config.text_config.head_dim = 16
config.text_config.max_position_embeddings = 512

config.vision_config.num_hidden_layers = 2
config.vision_config.hidden_size = 64
config.vision_config.intermediate_size = 128
config.vision_config.num_attention_heads = 4
config.vision_config.head_dim = 16
config.vision_config.image_size = 56

for subconfig in (config, config.text_config, config.vision_config):
    subconfig.dtype = "float32"
    subconfig.torch_dtype = "float32"

model = AutoModelForImageTextToText.from_config(config)
tokenizer = AutoTokenizer.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)
processor.image_processor.size = {"longest_edge": 56}

output_dir = "./tiny-random-mistral3"
os.makedirs(output_dir, exist_ok=True)
model.save_pretrained(output_dir, safe_serialization=True)
tokenizer.save_pretrained(output_dir)
processor.save_pretrained(output_dir)
```