Δw = η · xᵢ · yⱼ · J(p̂,p) = 1 − |p̂∩p| / |p̂∪p|

UnikAI

We build AI that learns, not just retrieves.

We're a small lab. Hydrogen-2 is our current model. 95.90% on GSM8K. BIM is our main research direction: a system that starts blank and updates its weights from interaction and corrections in real time, on a CPU. WebAI is our experiment making model training run entirely in a browser. Privy is live today.

See modelsResearch
Privy

Live · try it now

95.90%

GSM8K · stable model

CPU-native

Blank slate · learns in real time

What we work on

Biological learning, browser-native AI, a routing system, and one live product.

BIM

BIM 2 starts with zero knowledge and learns from interaction. When you correct it, it updates its weights. Restart it and it still knows. Transformers can't do that. They're frozen after training. BIM 2 is complete.

BIM docs →

WebAI

WebLM-1 is 30M parameters. It trains and runs entirely in a browser tab: TensorFlow.js, WebGPU, streaming generation, model export. An experiment, not a product.

WebAI →

MAUL

One idea: a model that only does math, one that only does code, one that searches — coordinated — outperform a single model doing everything. MAUL-2 with Hydrogen-2: 95.90% on GSM8K.

MAUL docs →

Privy

Describe what you want an assistant to do, set the rules, add files, chat. The only thing you can use today.

Privy docs →

Models & software

Hydrogen-2 is our current stable model. MAUL-2 routes it. BIM 2 is our active research.

Stable

Hydrogen-2

Current flagship model

95.90% on GSM8K. 97.56% on HumanEval. Paired with MAUL-2: math routes through SymPy, code through a sandbox. Same tests we run before calling a build stable.

Details →
Stable

MAUL-2

Orchestration system

Reads what you asked, picks a path (calculator, search, or direct answer), calls the right specialist, returns one reply. The hypothesis: specialization beats generalization.

Details →
Research

BIM 2

Biological learning system

131,072 cells. 64-sparse SDRs. Learns from interaction, updates weights in real time, runs on a CPU under 200MB. Complete implementation with benchmarks.

Details →

Scores

Hydrogen-2 with MAUL-2. The same tests we run before calling a build stable.

Bars are the score out of 100. From our last stable run.

GSM8K (math)95.90%

Hydrogen-2 with MAUL-2, same setup we ship as stable.

HumanEval (code)97.56%

Share of coding problems solved in one try.

Want to try something?

Privy is live. Open it if you want a working assistant now. If you're more curious about the research, BIM is the place to start.