Bijan Bowen tests Ant Group's Ling 3.0 Tiny, a 7.9-billion-parameter mixture-of-experts model that activates about 1.3 billion parameters per token and supports a 256,000-token context window. The text-only model responds very quickly and is expected to receive an open-weight release.
Ling produces a useful browser-style operating system, simple desktop interfaces and a surprisingly functional three-dimensional printer scene. It also handles roleplay with energy and creates a respectable keyboard product page, showing how much practical output a small active parameter count can deliver.
Harder tests expose clear limits. A C++ skateboard project requires a rewrite, the model cannot sustain the requested frame rate in a subway scene, and both a flight simulation and a V8-engine task fail. Bowen concludes that the model is promising for its size and speed, but its results are still too inconsistent for demanding work without close review.
Watch the original on YouTube