Marimba: a frozen 0.6B ASR backbone beats 30B Qwen3-Omni on 6 of 9 audio tests

solanotodesch · reddit · 2026-11-16

Author released the Marimba paper (arXiv:2611.18330): a "System One" speech model built on a frozen Qwen3-ASR-0.6B encoder, answering many typed questions about an 8s audio window from a single forward pass with calibrated probabilities.

Related event: Marimba: Open-Source 0.6B Speech Decision Model Beats 30B Rivals(2 posts)→

Original post →

More from Multimodal

Multimodal channel →