Running Liquid AI’s New Model on IBM AIX (No GPU Required)

Forget the H100 clusters for a moment. At SIXE, we decided to push enterprise hardware to its absolute limits to answer a burning question: Can a 2018-era IBM Power System, running AIX and relying purely on CPU, handle the latest generation of AI models?

We took Liquid AI’s new LFM2.5-1.2B model and ran it on an IBM POWER9 processor. To our knowledge, this is the first time an LFM2.5 model has ever run on AIX in Big-Endian mode.

The Hardware: IBM Power System S924

We used the workhorse of the enterprise world. Below are the specific configurations used for this benchmark:

Performance: Finding the Sweet Spot

We found that throwing every core at the model actually hurts performance due to synchronization overhead. The verdict was clear: Using just 8 cores in SMT-2 mode (16 threads) gave us 26.7 tokens per second.

This efficiency is possible because LFM2.5 is a hybrid architecture designed for extreme efficiency, mixing Convolutional blocks (shortconv) for speed and Attention layers (GQA) for context.

Real-World Test: The SysAdmin Gauntlet

Numbers are nice, but can it actually work? To prove this isn’t just a benchmark toy, we put LFM2.5 through real AIX administrative tasks and compared it against a standard Transformer (TinyLlama 1.1B).

Round 1: The Cryptic Error (errpt)

We fed the models a raw AIX error log regarding a power supply failure.

• ❌ TinyLlama 1.1B: Fails. It got stuck in an infinite loop repeating “PWRSPLY”.
• ✅ LFM2.5 1.2B: Pass. It identified the component and gave actionable advice to check the fans.

Round 2: The Security Audit (last)

We provided a login history log containing a hidden crash event.

• ❌ TinyLlama 1.1B: Fails. Absolute silence; it generated one token and stopped.
• ✅ LFM2.5 1.2B: Pass. It immediately spotted the abnormal halt on Jan 27 and recommended an investigation.

Round 3: The Dangerous Advice (/etc/passwd)

We asked the models to audit a standard password file. The results here were shocking.

• ❌ TinyLlama 1.1B: CATASTROPHIC FAIL. It claimed the “root” user was not required and recommended deleting it. Following this advice would destroy the server.
• ✅ LFM2.5 1.2B: Pass. It correctly identified actual potential risks like “guest” and “nobody” accounts with high UIDs.

Round 4: Service Hardening (lssrc -a)

We asked the models to review running services and recommend hardening steps.

• ❌ TinyLlama 1.1B: Fails. Silence again.
• ✅ LFM2.5 1.2B: Pass. It flagged risky services like sendmail and portmap, and provided the correct AIX command (stopsrc) to disable them.

Why This Matters for IBM Power Users

This benchmark proves that IBM Power Systems are capable AI inference engines for critical, on-premise tasks:

• Data Sovereignty: Analyze sensitive errpt logs, financial data, or user audits locally. No data leaves your server.
• Legacy Modernization: Use local LLMs to help understand and document legacy COBOL or C code residing on the server.
• Efficiency: You don’t need a GPU cluster. You likely already own the hardware capable of doing this.

Try It Yourself

We believe in open source. We have released the AIX port and the converted Big-Endian models.

Code: gitlab.com/librepower/llama-aix
Models: huggingface.co/librepowerai