title: README
emoji: 🚀
colorFrom: gray
colorTo: purple sdk: gradio pinned: false

🚀 Big-Endian Models for IBM AIX and IBM i

Welcome! This organization hosts machine learning models adapted and validated for Big Endian architectures, with a primary focus on:

• IBM AIX on Power Systems (now with 10 LLMs ready to use)
• CPU-only inference (no GPU required)
• Real enterprise environments (not toy benchmarks)

Our goal is simple:

Make modern AI usable on legendary mission-critical platforms.

No vendor lock-in. No mandatory accelerators. Just practical inference where your workloads already live.

📦 Available Models (v0.1.1 — Production Ready)

10 big-endian GGUF models, fully tested on AIX 7.3 POWER9:

Benchmarks: AIX 7.3 TL4, IBM POWER9 @ 2.75 GHz, 16 threads (SMT-2), GCC 13.3.0

🚀 Quick Start

Download a model and run on AIX:

# Clone and build
git clone https://gitlab.com/librepower/llama-aix.git
cd llama-aix
./scripts/build_aix_73.sh

# Get a model (example: fastest model)
mkdir -p models
wget https://huggingface.co/librepowerai/LFM2.5-1.2B-Instruct-Q4_K_M-BE/resolve/main/LFM2.5-1.2B-Instruct-Q4_K_M.gguf -O models/

# Run inference
export LIBPATH=$PWD/build/bin
./build/bin/llama-simple -m models/LFM2.5-1.2B-Instruct-Q4_K_M.gguf -n 256 "Your prompt"

All models available at: https://huggingface.co/librepowerai

---
🔬 What you will find here

Models and artifacts specifically prepared for:

- Big Endian compatibility
- AIX and IBM i runtime environments
- CPU inference (optimized with VSX SIMD + OpenBLAS)
- Minimal external dependencies

v0.1.1 Optimizations:
- VSX auto-vectorization: -O3 -mvsx -maltivec (+6.7% performance)
- OpenBLAS BLAS backend: GEMM acceleration for attention layers
- GCC 13.3.0 native build: no xlc required

Typical contents:

- Big-Endian converted GGUF models (Q4_K_M quantization)
- BE-safe tokenizer assets
- Complete build scripts and documentation
- Performance benchmarks on real Power hardware

Everything here has been tested on actual AIX systems — not emulators.

---
⚙️  Tooling

Models here are intended to run with:

- llama-aix - Full port of llama.cpp to AIX 7.3
https://gitlab.com/librepower/llama-aix
- Native AIX / IBM i toolchains (GCC or IBM Open XL)
- CPU inference (no GPU acceleration needed)

We deliberately avoid GPU assumptions and licensing complexity.

---
🧭 Why Big Endian?

Most open-source AI models today assume:

- Little Endian
- Linux
- x86 or CUDA

Enterprise reality is different.

Many production systems run on:

- IBM AIX or IBM i
- POWER9 / POWER10 / POWER11
- Big Endian binaries
- Highly regulated, long-lifecycle environments

Porting models to BE is not trivial:
tokenizers, SIMD paths, memory layout, third-party deps — everything matters.

This repository exists to close that gap.

---
🌱 Philosophy

We believe innovation is often not about new hardware —
it is about unlocking what you already own.

That means:

- reducing infrastructure footprint
- maximizing TCO
- extending the life of enterprise platforms
- keeping data on-prem when needed
- enabling AI without massive architectural changes

This is practical engineering work, driven by real use cases — not marketing slides.

---
📚 Documentation & Research

- Build Guide: https://gitlab.com/librepower/llama-aix/-/blob/main/docs/BUILD_AIX_73.md
- Troubleshooting: https://gitlab.com/librepower/llama-aix/-/blob/main/docs/TROUBLESHOOTING_AIX.md
- VSX Research: https://gitlab.com/librepower/llama-aix/-/blob/main/docs/VSX_OPTIMIZATION_RESEARCH.md (Phase A & B complete, Phase C roadmap for v0.2)
- Blog: https://sixe.eu/news/running-liquid-ais-new-model-on-ibm-aix-no-gpu-required

---
🔗 Related projects

- LibrePower Open Source Initiative
https://librepower.org
- llama-aix (this port)
https://gitlab.com/librepower/llama-aix
- AIX Ports (RPM packages)
https://gitlab.com/librepower/aix
- llama-ibm-i
https://ajshedivy.notion.site/How-to-run-LLMs-on-IBM-i-1ce662038dd180f8b59bd9cfada2815b

---
✉️  Contact

hello {at} librepower.org

#LibrePower — Unlocking Power Systems through open source
Minimal footprint. Unmatched RAS. Better TCO.

Last updated: February 18, 2026 — v0.1.1 production release with 10 models