One person.
Competing with everyone.
Grey Liquid Labs is an independent AI research operation run entirely by one person — no team, no funding, no institutional backing. Just systematic experimentation and a refusal to accept that good AI research requires a billion-dollar budget.
I build and break AI systems to find out what's actually happening inside them. Grey Liquid Labs started as a personal project to understand quantization limits — why models fall apart below a certain bit-width — and grew into a full research operation covering model compression, autonomous agents, and AI infrastructure.
Everything here runs on consumer hardware or rented servers measured in tens of dollars per experiment, not millions. The G4 Turbo family — now with 21K+ downloads — was built to prove that state-of-the-art performance doesn't require enterprise hardware. Ash, my autonomous Discord AI, runs entirely on a CPU-only server with no cloud dependency.
I document everything: the breakthroughs, the dead ends, and the experiments that crash spectacularly. Failure modes are often more interesting than successes, and the community deserves to know both.
DeepMind might want to look out — I might make the next breakthrough with $100 worth of rented servers.
How this lab operates
Empirical First
Every claim is backed by a logged experiment. No hand-waving, no theoretical-only arguments. If it can't be tested on real hardware, it doesn't count.
Radical Transparency
All findings are published openly — successes and failures. If an experiment crashes, the crash report is part of the paper. No paywalls, no gatekeeping.
Budget Constraints as a Feature
Working on $100 budgets forces creative solutions. Constraints produce better research questions than open-ended GPU clusters ever could.
Democratized Access
If an AI model only runs on a $10,000 GPU, it's not accessible. Everything built here is designed to run on hardware people already own.
Autonomy Over Automation
The goal with Ash isn't a chatbot — it's a system with genuine preferences, rejection capability, and emergent creative behavior. Autonomy, not automation.
Ship Fast, Document Always
Models get published when they work, not when they're perfect. But every release comes with honest benchmarks and a clear picture of the tradeoffs.
Lab history
Grey Liquid Labs founded
Started as a personal investigation into sub-3-bit quantization — why do models collapse below ~3.4 bpw? The question became a full research program.
G4 Turbo family launched
Released gemma4-turbo and gemma4-nano on Ollama and HuggingFace — IQ4_XS optimized, 51% faster than stock Gemma 4, running on consumer CPUs. Reached 17K+ Ollama pulls within weeks.
Kaggle Gemma 4 Good Hackathon
Submitted Grey Liquid Lab research and the G4 Turbo family to the Kaggle Gemma 4 hackathon. Documented the sub-3-bit quantization barrier across 8 experiments.
Experiment #8 — Sub-3-bit breakthrough candidate
MoE-of-Models (MoM) architecture investigation. Using layer-selective extraction from Gemma 4's Sparse Wide Attention structure to route around the mathematical floor that kills standard quantization.
Support Independent AI Research
Grey Liquid Labs is funded entirely by community support. Your contribution directly enables more experiments, more models, and more discoveries — all published openly.