Edge inference
Running capable models on the device, not the data centre. We are looking for alternatives to pruning and quantisation that keep what the model can do.
Ongoing
Edge inference / ongoing
Research on running capable models locally.
Arithemiq is an independent research lab in Bengaluru. We work on new ways to fit language and vision models on small hardware, beyond pruning, quantisation and low-rank adapters.
Research
Adaptation
Specialising models without LoRA or QLoRA style low-rank adapters, at a cost a small device can bear.
Ongoing
Over time: language, vision and video models, including spatial and 3D.
Experiments
Built in-house. None are released yet.
-
X.01
ParseSense
A data extraction and sync pipeline. Pulls web data into databases, CRMs and AI ingestion, with scheduled ETL and ELT.
Built / open source planned
-
X.02
Learn-Gen-X
Generates explainer videos shaped to teach.
Prototype
-
X.03
Settle-X
Financial planning and cash-flow modelling, with forward-looking recommendations.
Prototype
-
X.04
Voice-X
A sequence-to-sequence speech model that works from voice to voice, without a separate transcribe and synthesise step.
In progress
-
X.0…
Next
More in progress.
Ongoing