Instructions to use formulae/mita-v1.2-7b-2-24-2025 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use formulae/mita-v1.2-7b-2-24-2025 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="formulae/mita-v1.2-7b-2-24-2025")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("formulae/mita-v1.2-7b-2-24-2025")
model = AutoModelForCausalLM.from_pretrained("formulae/mita-v1.2-7b-2-24-2025")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use formulae/mita-v1.2-7b-2-24-2025 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "formulae/mita-v1.2-7b-2-24-2025"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "formulae/mita-v1.2-7b-2-24-2025",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/formulae/mita-v1.2-7b-2-24-2025

SGLang

How to use formulae/mita-v1.2-7b-2-24-2025 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "formulae/mita-v1.2-7b-2-24-2025" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "formulae/mita-v1.2-7b-2-24-2025",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "formulae/mita-v1.2-7b-2-24-2025" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "formulae/mita-v1.2-7b-2-24-2025",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use formulae/mita-v1.2-7b-2-24-2025 with Docker Model Runner:
```
docker model run hf.co/formulae/mita-v1.2-7b-2-24-2025
```

Formulae/MITA-V1.2-7B-2-24-2025

Overview

Formulae/MITA-V1.2-7B is designed primarily for mathematics, with some capability in coding. Built using the Linear DARE merge method, this model blends powerful mathematical reasoning with computational accuracy.

Rank	Type	Model	Average	IFEval	BBH	MATH	GPQA	MUSR	MMLU-PRO	CO₂ Cost
914	🤝	formulae/mita-v1.1-7b-2-24-2025	29.48 %	34.12 %	35.44 %	43.50 %	8.61 %	16.06 %	39.15 %	0.67 kg
1403	🤝	formulae/mita-v1.2-7b-2-24-2025	24.86 %	25.64 %	28.41 %	48.79 %	7.49 %	12.63 %	26.21 %	0.64 kg

Merge Details

Base Model: Qwen/Qwen2.5-Math-7B-Instruct
Merged Models:
- nvidia/AceMath-7B-Instruct
- open-r1/OpenR1-Qwen-7B
Merge Method: Linear DARE
Data Type: bfloat16
Merge Parameters:
- Density & Weight: 0.5 for AceMath & OpenR1
- Normalization: Disabled
- Int8 Masking: Enabled

What is DARE?

DARE (Density-Aware Residual Estimation) is an advanced model merging technique designed to preserve task-specific knowledge. Unlike simple model averaging, DARE adjusts parameter density to ensure that merged models retain their specialized skills while improving general performance.

📖 Reference: DARE Paper

This merge is also inspired by task arithmetic, which shows that models can be linearly combined to enhance capabilities in specialized domains.

📖 Reference: Task Arithmetic Paper

Capabilities

✅ Advanced Mathematics – Strong problem-solving, algebra, calculus, and theorem applications.
✅ Limited Coding Support – Can handle basic programming tasks but is not optimized for complex software development.

Limitations & Risks

⚠ Hallucinations in Code – Not a coding-specialized model, may produce incorrect or insecure implementations.
⚠ Arithmetic Errors – While highly capable, the model still makes occasional miscalculations.

Usage Disclaimer

Formulae/MITA-V1.2-7B is an experimental mathematical model. For verified accuracy, always cross-check results with reliable tools.

Contribute

We welcome contributions, including quantizations, fine-tuning, and further enhancements.

💡 Support Us: Buy Me a Coffee

📩 Contact: formulaeresearch@gmail.com

Future Development

This is part of the MITA series. Future iterations will integrate MoE (Mixture of Experts) for even more specialized reasoning across multiple domains.

Made possible with MergeKit.