# Underdog releases a 163 MB speech model built from NVIDIA's Parakeet > Source: > Published: 2026-10-21 05:16:08+00:00 # Underdog releases a 163 MB speech model built from NVIDIA's Parakeet **Founder Sigil Wen says Bark 3.0 transcribes meeting audio locally; its published benchmark puts it level with Parakeet on average, not ahead.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) · Published Primary source: [X](https://x.com/UnderdogAI/status/2112733007484518470) ## Why it matters Bark makes Conway's on-device AI thesis concrete: a 163 MB speech model with competitive company-reported accuracy, built from existing models and published under a permissive license. Its benchmark table supports a strong result, not an unqualified claim of being the best. [Sigil Wen](https://sigilwen.ca/index.html), founder of Conway Research, released [Underdog Bark 3.0](https://huggingface.co/ConwayResearch/Underdog-Bark-3.0) on October 21st, pitching a 163 MB [speech-to-text](https://runtimewire.com/models/huggingface/codenamewei-speech-to-text-0e73183c7962a8ad) model for laptops and phones. In his [announcement on X](https://x.com/UnderdogAI/status/2112733007484518470), Wen said it matches or beats larger systems on recorded-meeting transcription while running locally. The model weights are available under the CC-BY 4.0 license.… When you purchase through links in our articles, [we may earn a small commission](https://runtimewire.com/affiliate-monetization-standards). This doesn’t affect our editorial independence.