wpferrell/llama-3.1-8b-instruct-bigsmall

Llama 3.1 8B Instruct (BigSmall compressed)

15.0 GB -> 9.75 GB (BF16). Lossless. Zero inference overhead. Any hardware.

Compressed with BigSmall -- decompresses once at load time, then runs at full native speed. Every weight is bit-identical to the original.

Why BigSmall

vs quantization (llama.cpp, GGUF, AWQ, bitsandbytes)

Quantization permanently degrades weights. BigSmall is lossless -- bit-identical weights, no accuracy loss, fine-tuning safe, fully reproducible.

vs DFloat11 (runtime lossless compression)

DFloat11 keeps weights compressed during inference -- saves VRAM but adds ~2x overhead at batch=1, CUDA only. BigSmall decompresses once at load time and runs at full native speed on any hardware.

	BigSmall	DFloat11
Compression ratio (BF16)	65-66%	~70%
Inference overhead	None	~2x at batch=1
Hardware	CPU, Apple Silicon, AMD, any GPU	CUDA only
FP32 / FP16 / FP8 support	Yes	BF16 only
Fine-tuning safe	Yes	No
Streaming loader (< 2GB RAM)	Yes	No

vs ZipNN (storage lossless compression)

Same category as BigSmall -- decompresses at load time. BigSmall compresses better (65% vs 67% BF16) and supports more formats. BigSmall also has a streaming loader so you can run 70B models with under 2GB peak RAM.

Install

ash pip install bigsmall

Load

`python import bigsmall bigsmall.install_hook()

from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("wpferrell/llama-3.1-8b-instruct-bigsmall") `

Stream layer by layer (peak RAM under 2GB even for 7B models)

`python from bigsmall import StreamingLoader from transformers import AutoModelForCausalLM

with StreamingLoader("wpferrell/llama-3.1-8b-instruct-bigsmall", device="cuda") as loader: model = loader.load_model(AutoModelForCausalLM) `

Compression stats

Original	Compressed	Ratio	Format	Verified
15.0 GB	9.75 GB	65.0%	BF16	md5 every tensor

• GitHub: wpferrell/Bigsmall
• PyPI: pip install bigsmall
• All pre-compressed models: huggingface.co/wpferrell