Tether Builds 4 Billion-Parameter AI Models for Phones as Scaling Laws Face Diminishing Returns
Updated
Updated · Computerworld · Oct 27
Tether Builds 4 Billion-Parameter AI Models for Phones as Scaling Laws Face Diminishing Returns
1 articles · Updated · Computerworld · Oct 27
Summary
Tether’s compact AI lineup tops out at about 4 billion parameters, and some models can run on a basic smartphone offline after download.
That push targets a growing backlash against ever-larger AI systems, as trillion-parameter models demand huge compute budgets while gains from brute-force scaling increasingly taper off.
QVAC’s latest Genesis III dataset expands synthetic training data to 191 billion tokens across 19 STEM domains, aiming to boost specialized performance through curated data rather than sheer size.
The company pairs that with quantization and SDK tools including TurboQuant—cutting KV-cache memory by up to 80%—plus TurboVec and clustered inference for heavier on-device workloads.
Tether says the approach lets specialized models in areas such as medical reasoning, translation and vision outperform larger general-purpose systems on specific tasks using consumer hardware.