Infinity Intelligence
The part that continues
Labs is the research subdivision of Infinity Intelligence: experimental architectures, novel training techniques, and the model programs that carry on now that frontier development has concluded.
What Labs is
For most of the division’s life, Labs was the research function attached to a frontier model program. It explored architectures, efficiency techniques and reasoning systems, and the results fed into the Xora family.
With the frontier program concluded, that relationship inverts. Labs is now the centre of the division rather than a function of it, and the two remaining model programs — PolyCode and the WillMe GPT line — are run out of it.
The practical difference is scale, not ambition. Labs works on models small enough to be sustainable, in the open where possible, and publishes what it finds.
Research areas
Efficiency
Getting more capability out of less hardware. Quantization, sparsity, and memory-efficient training and inference, aimed at models that run on infrastructure a customer already owns.
Architectures
Alternatives to dense attention, including the linear recurrent models behind WillMe GPT and the diffusion approach behind PolyCode.
Context scaling
Making long context affordable rather than merely possible. Most of this work is evaluated in the open, on the WillMe AI platform.
Reasoning
How models allocate effort. Four generations of Xora work went into deciding when a model should think longer, and that research continues independently of the model family that funded it.
Active programs
Two model programs continue under Labs.
PolyCode
A small diffusion-based coding model that denoises a whole syntax tree rather than predicting a program token by token. Built for local and air-gapped deployment, where the Xora family found its users.
Model card →WillMe GPT
Efficient linear recurrent models used to test context scaling and training techniques with real users, on the public WillMe AI platform rather than behind a research paper.
Model card →