0.5B Rumble-2-Flash ASR beats Whisper large-v3 with 30% fewer errors at a third the size
ricklamers · x · 2026-11-18
ricklamers highlights Rumble-2-Flash (0.5B, Apache 2.0, 45 languages): 67% smaller than Whisper large-v3 (1.02 vs 3.09 GB) with 30% fewer errors (5% vs 7.2% WER), and 31% fewer errors than Whisper turbo (5.9% vs 8.6%); on AMI meetings it cuts WER from 16% to 9%. He also one-shotted RumbleBar, an MIT-licensed macOS menu bar dictation app running the model locally on Apple Silicon via Metal — press ⌥D to dictate, transcript pastes into any app with audio never leaving the device.
More from Infra
- Wall Street Execs and NSA Tell Congress: AI-Discovered Vulnerabilities Triggered Outage-Causing Patching Rush — Jsevillamol · 2026-11-18
- Nvidia halts shipments of US-approved AI chips to China, Reuters reports — minchoi · 2026-11-18
- Analyst stands by $20 trillion Nvidia market cap thesis by 2030 — Beth_Kindig · 2026-11-18
- armless: A Rust-Written AArch64 Emulator Boots Linux 6.8 in a Browser Tab — sull · 2026-11-18
- Named imports fix V8 inlining: OTel JS Tracer gets ~2x throughput — DanielLockyer · 2026-11-18
- Fast path for primitive attributes boosts OTel span creation up to 24% — DanielLockyer · 2026-11-18