Taco

Taco: LLM compression software that produces identical compressed bytes on every CPU and GPU.

Taco is a lossless compressor utilizing the GPT-2 language model to predict data and a range coder to store unpredicted portions. Its inference engine, HardShell, performs calculations exclusively using integer arithmetic to ensure every CPU and GPU calculates identical predictions bit for bit. The software includes an exact converter, multiple backend implementations, and conformance vectors connecting the model to real file bytes. It supports CPU implementations with NEON and AVX2, along with NVIDIA GPUs via CUDA and other GPUs through wgpu supporting Vulkan, Metal, and DX12. Designed as an experimental alpha proof of concept, it targets developers and researchers needing cross-platform compression consistency.

Taco at a glance

Best for: Software engineer
Runs on: macOS, Windows, Linux, CLI
Pricing: Free
Released: October 2026
Category: AI coding tool
Website: github.com/tac0mania/taco

Key features

Lossless data compression using a GPT-2 language model
Integer-only arithmetic for cross-machine bit consistency
CPU implementations supporting NEON and AVX2
GPU implementations supporting CUDA, Vulkan, Metal, and DX12
Exact converter for consistent model file generation
Conformance vectors connecting model weights to file bytes
Command-line interface for folding and unfolding files

Use it for

Compressing text files with cross-machine compatibility checks
Testing cross-platform model inference consistency
Experimenting with language model entropy coding

Alternatives to Taco

AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
LMDeploy — Accelerates inference with high-performance engines and quantization tools.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
Vramkit — It cuts GPU memory usage using verified benchmarks on real models.
fal — Access 1,000+ generative models and high-performance serverless GPUs.
whichllm — Ranks models using real benchmarks and hardware fit instead of parameter count.
llm-d — Deploys production-grade LLM serving on existing hardware clusters.
Soup — Fine-tune and align models like Llama-3.1-8B on a 4 GB laptop GPU.
Enki LLM — Runs Llama-3.2 1B in pure Rust at 40 tok/s with live GPU to CPU migration.
Neural Compressor — Supports advanced quantization and sparsity for large language models.
MTPLX — It runs local models twice as fast via native MTP speculative decoding.
GPTCache — It decreases expenses and enhances performance by caching LLM query results.

All Taco alternatives