The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Supporting case

An Alibaba team publishes weights for a 173 MB model that can run locally

What happened

The Taobao MNN team’s November 9 technical paper ↗ describes a model with roughly 300 million parameters intended to run on a compatible device. Its public repository provides downloadable weights under Apache 2.0. The four-bit build takes about 173 MB in total and runs through the team’s MNN software. The exact date when the weights first became public is unknown. The team recommends the model for narrow tasks such as extracting fields from text and cautions against using it for difficult reasoning or extended coding. Performance figures are the developer’s own tests.

A small institution could run a suitable task on its own device without sending each request to a hosted model. The permissive license and small download make that option more practical for narrow tasks. Compatibility and the quality of local results still limit which workflows can move.

Dates and implementation stage

November 9, 2026: technical paper dated. The exact date the weights first became public is unknown.

Supporting sources

Related trends