README.md

---
license: apache-2.0
tags:
  - bigsmall
  - compression
  - lossless
  - qwen
  - qwen2.5
---

# Qwen 2.5 7B Instruct (BigSmall compressed)

**15.2 GB -> 10.05 GB. Full quality - not quantization.**

This is Qwen2.5-7B-Instruct compressed with [BigSmall](https://github.com/wpferrell/Bigsmall) - lossless neural network weight compression. Every weight is bit-identical to the original. No accuracy loss whatsoever.

## Install

`ash
pip install bigsmall
`

## Load and run inference (streaming)

`python
from bigsmall import StreamingLoader
from transformers import AutoModelForCausalLM, AutoTokenizer

# Streams one layer at a time - saves significant peak RAM
loader = StreamingLoader("wpferrell/qwen2.5-7b-instruct-bigsmall")
model = loader.load_model(AutoModelForCausalLM)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

messages = [{"role": "user", "content": "Explain lossless compression in one paragraph."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))
`

## Or decompress to disk first

`python
from bigsmall import from_pretrained
from transformers import AutoModelForCausalLM
model = from_pretrained("wpferrell/qwen2.5-7b-instruct-bigsmall", model_class=AutoModelForCausalLM)
`

## Or use the AutoModel hook (transparent)

`python
import bigsmall
bigsmall.install_hook()

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("wpferrell/qwen2.5-7b-instruct-bigsmall")
`

## Compression stats

| Metric | Value |
|--------|-------|
| Original size | 15.2 GB |
| Compressed size | 10.05 GB |
| Ratio | 66.0% (BF16) |
| Format | BF16 -> BigSmall (.bs shards) |
| Lossless verified | md5 every tensor |

## Comparison

| Tool | BF16 Ratio | FP32 Ratio | Inference Overhead | Hardware |
|------|------------|------------|-------------------|---------|
| [ZipNN](https://arxiv.org/abs/2411.05239) | 67% | 83% | None | CPU |
| [DFloat11](https://arxiv.org/abs/2504.11651) | ~70% | BF16 only | ~2x at batch=1 | CUDA only |
| [ZipServ](https://arxiv.org/abs/2603.17435) | ~70% | BF16 only | 1.22x faster | GDDR GPU |
| **BigSmall** | **66.0%** | **75.5%** | **None** | **CPU + any GPU** |

*Lower ratio = better compression.*

## About BigSmall

BigSmall compresses at the joint entropy floor for neural network weights. It codes sign+exponent jointly and mantissa conditioned on exponent, achieving the information-theoretic minimum. The streaming loader decompresses one transformer layer at a time directly into VRAM.

- GitHub: [wpferrell/Bigsmall](https://github.com/wpferrell/Bigsmall)
- PyPI: `pip install bigsmall`