AnalysisAI ModelsNovember 3, 2026

Blog disputes Astra's 'read and write weights' Qwen claim

Read original source →honnibal.dev

The critique says only experiments 1, 2 and 4 used raw weights alone, on networks of 17K–172K weights; every real language-model experiment used Python programs. The headline Qwen3.5-0.8B result also ran the model on up to 50 batches of practice problems.

1 source

More stories today

Open the live feed