README.md

---
license: apache-2.0
tags:
  - nebula-s
  - svms
  - math-reasoning
  - competition-math
  - 4bit
  - quantized
  - bitsandbytes
library_name: transformers
---

# Nebula-S-v1-4bit

4-bit quantized version of [Nebula-S-v1](https://huggingface.co/punitdecomp/Nebula-S-v1).

**Nebula-S-v1** is a reasoning-enhanced language model using the **SVMS (Structured-Vector Multi-Stream)** architecture.

## What's different from Nebula-S-v1?

| | Nebula-S-v1 | Nebula-S-v1-4bit |
|---|---|---|
| Backbone precision | bf16 | **4-bit (nf4)** |
| Adapter precision | bf16 | bf16 |
| Backbone size | ~8 GB | **~2 GB** |
| Total size | ~9 GB | **~3 GB** |
| VRAM needed | ~18 GB | **~6 GB** |
| Requires | CUDA / MPS / CPU | **CUDA only** (bitsandbytes) |

## Quick Start

```bash
pip install torch transformers>=4.51.0 bitsandbytes accelerate huggingface-hub
```

### Option 1: Using huggingface_hub

```python
from huggingface_hub import snapshot_download
import sys

snapshot_download("punitdecomp/Nebula-S-v1-4bit", local_dir="./Nebula-S-v1-4bit")
sys.path.insert(0, "./Nebula-S-v1-4bit")
from nebula_s import load_nebula_s

model, tokenizer = load_nebula_s("./Nebula-S-v1-4bit", device="cuda")
```

### Option 2: Using git clone

```bash
git lfs install
git clone https://huggingface.co/punitdecomp/Nebula-S-v1-4bit
```

```python
import sys
sys.path.insert(0, "./Nebula-S-v1-4bit")
from nebula_s import load_nebula_s

model, tokenizer = load_nebula_s("./Nebula-S-v1-4bit", device="cuda")
```

### Generate a response

```python
messages = [{"role": "user", "content": "Solve step by step: what is 17 * 23?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
response = model.generate(
    inputs["input_ids"], inputs["attention_mask"],
    tokenizer, max_new_tokens=2048, temperature=0.7
)
print(response)
```

## License

Apache 2.0. Backbone derived from an Apache-2.0 licensed base model.