README.md

---
license: apache-2.0
tags:
  - nebula-s
  - svms
  - math-reasoning
  - competition-math
  - quantized
  - int4
  - hqq
library_name: transformers
---

# Nebula-S-v1-lite

Lightweight (~3GB) version of [Nebula-S-v1](https://huggingface.co/punitdecomp/Nebula-S-v1), pre-quantized to int4 using [HQQ](https://github.com/mobiusml/hqq) (Half-Quadratic Quantization).

**Runs on Mac (MPS), CUDA, and CPU.**

| Variant | Download | Runtime | Platform |
|---|---|---|---|
| [Nebula-S-v1](https://huggingface.co/punitdecomp/Nebula-S-v1) | ~9 GB | ~9 GB | Universal (bf16) |
| [Nebula-S-v1-4bit](https://huggingface.co/punitdecomp/Nebula-S-v1-4bit) | ~3 GB | ~3 GB | CUDA only (bnb) |
| **Nebula-S-v1-lite** | **~3 GB** | **~3 GB** | **Mac + CUDA + CPU** |

## Quick Start

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

### Option 1: Using huggingface_hub

```python
from huggingface_hub import snapshot_download
import sys

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

# Auto-detects device (mps on Mac, cuda on NVIDIA, cpu fallback)
model, tokenizer = load_nebula_s("./Nebula-S-v1-lite")
```

### Option 2: Using git clone

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

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

model, tokenizer = load_nebula_s("./Nebula-S-v1-lite")
```

### 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)

device = next(model.parameters()).device
inputs = tokenizer(text, return_tensors="pt").to(device)

response = model.generate(
    inputs["input_ids"], inputs["attention_mask"],
    tokenizer, max_new_tokens=1024, temperature=0.7
)
print(response)
```

### Explicit device

```python
# Mac
model, tokenizer = load_nebula_s("./Nebula-S-v1-lite", device="mps")

# NVIDIA GPU
model, tokenizer = load_nebula_s("./Nebula-S-v1-lite", device="cuda")

# CPU
model, tokenizer = load_nebula_s("./Nebula-S-v1-lite", device="cpu")
```

## License

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