Bookmarks

    Nothing saved yet. Use the bookmark on any article. Bookmarks are kept on this device only: they stay private and work without an account.

    Settings

    Appearance

    Reading mode

    Font size

    Content width

    Line spacing

    Keyboard shortcuts

    /
    Search
    j k
    Next / previous story
    ← →
    Previous / next section (columns)
    b
    Bookmarks
    s
    Settings
    r
    Columns / scroll
    t
    Light / dark

    Saved in this browser only.

    Ping2Prompt
    Subscribe to the Weekly Brief Sign in
    Purple looping tubes and small gold cubes float around a Hugging Face logo with a smiling yellow and purple face on a grey ba
    Photo by Brecht Corbeel on Unsplash

    Rumble-2-Flash: Reson8's 0.5B open speech model for 40+ languages on Hugging Face

    Hugging Face

    Rumble-2-Flash weights are public under Apache 2.0

    Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model covering 40+ languages. The weights are on the Hugging Face Hub at Reson8/Rumble-2-Flash under the Apache 2.0 license, and the model is also offered through the vendor's API.

    Distilled from Rumble-2 on the Qwen3-ASR architecture

    The model uses the Qwen3-ASR architecture, so Transformers and vLLM support it from day 0, with no forked runtime and no trust_remote_code. It was distilled from Reson8's larger production model, Rumble-2, using on-policy distillation: the student transcribes audio itself, and the teacher grades the transcript token by token using its own probabilities. Qwen3-ASR's vocabulary came with the architecture, so teacher and student use different tokenizers. The authors handle this with GOLD, which merges tokens the two tokenizers split differently, compares shared tokens directly and compares the rest by sorted probabilities.

    Language detection and output controls

    The model detects the language automatically across 40+ languages. You can force a language by passing the language argument. Output options cover numbers as digits or words, fillers, verbatim repeats and a forced language.

    Benchmark figures are Reson8's own, not independent

    All figures below are the vendor's claims. The sources include no independent evaluation.

    • Multilingual: Reson8 reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track (Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, English), better than Meta's omniASR-LLM 7B (15x larger) and ahead of Whisper large-v3.
    • English: Reson8 reports a 6.6% average WER on the Open ASR Leaderboard English track.
    • Call centre: in a vendor-run test on AppTek's benchmark (302 hours of helpline calls, 14 English accents), Rumble-2-Flash scored 7.86% WER against 7.78% for NVIDIA Nemotron 3.5 Transcribe (30B), 8.7% for Qwen3-ASR 0.6B and 7.4% for Qwen3-ASR 1.7B.
    • Throughput: Reson8 reports more than 7,000 RTFx on the larger Open ASR Leaderboard datasets where batches stay full, about an hour of audio in half a second. Throughput depends on how full the batches are, so expect lower numbers with sparse traffic. Reson8 also says the model competes with models 3-60x its size.

    Running Rumble-2-Flash in Transformers

    For Transformers, install the dependencies with uv pip install torch transformers accelerate, then load the model with AutoProcessor and AutoModelForMultimodalLM. Audio input can be a local path, a URL, a NumPy array, or a list of these. The commands come from the documentation listed under Sources and should be checked against your own setup before running.

    Who should test Rumble-2-Flash

    If you self-host transcription, this is a small model with a permissive license that fits stacks you may already run on Transformers or vLLM. Test it on your own audio before switching, since the benchmark figures are vendor-reported and your accents, noise and languages may differ.

    What Reson8 has not published yet

    The sources give no API pricing, rate limits or hardware and memory requirements. They also do not list every supported language.

    Sources

    Image: Photo by Brecht Corbeel on Unsplash


    This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].

    Continue

    Share on Bluesky · Mastodon · Email ·

    Discussion