Phonon-2: a 164MB speech model beats Whisper, runs fully on-device

Fermion's 164MB ASR hits 5.21% WER on a laptop — plus its open Neutrino LLMs, a new multilingual TTS leaderboard, and source-aware MCP agents.

Nowline OCT 1 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Phonon-2 beats a model 10x its size

    Phonon-2 is a 164MB English ASR model (ternary-quantized to ~2.1 bits/weight) that posts 5.21% average WER across seven public test sets — beating Whisper large-v3-turbo, which weighs 1,618MB. Every open model that scores higher is at least 5.8x its size. It's CC-BY-4.0 and transcribes an hour of audio in ~20s on a MacBook Air GPU (174x realtime), ~25s on CPU, so you can ship fully offline dictation and transcription with zero API cost.

  • Fermion also dropped the open Neutrino LLMs

    Alongside the ASR model, Fermion open-sourced Neutrino-8B (Apache-2.0, built on Qwen3-8B, 40,960-token context, packed to 3.88GB), scoring 73.17% on IFEval prompt-strict and 67.84% on MMLU-Redux, plus a 0.6B-Chat drafter for speculative decoding that stays token-identical to 8B greedy at up to 1.92x faster. That's a compact, tool-calling local stack — big model plus matched drafter — you can self-host today.

  • A new Open TTS Leaderboard to pick your voice model

    Hugging Face's TTS-AGI launched the Open TTS Leaderboard: a scalable, reproducible benchmark for multilingual text-to-speech and voice cloning. If you've been guessing which open TTS to ship, there's finally a head-to-head board to check before you wire one into your product.

  • ProvenanceGuard: stop MCP agents citing the wrong source

    New work on source-aware verification for MCP-based agents (ProvenanceGuard) checks not just whether an answer is factually correct but whether it actually came from the source the agent cites. For anyone wiring tools into agents over MCP, it targets the exact failure that erodes trust — right fact, invented citation.

  • NVIDIA Kumo pushes the tabular-prediction frontier

    NVIDIA's Kumo Tabular claims a new accuracy-efficiency frontier for tabular prediction and ships as an NIM container with a public model card. If your product still runs on structured or spreadsheet data, it's a foundation-model option worth benchmarking against your gradient-boosted baselines.