A half-billion parameters is now a serious transcription model

The useful number in Reson8's announcement is not the parameter count. It is the comparison the count makes possible. Rumble-2-Flash averages 6.96% error across nine languages on the Open ASR Leaderboard's multilingual track. The announcement notes that this puts it ahead of Whisper large-v3 and ahead of Meta's omniASR-LLM 7B, a model roughly 15 times its size. On the English track it averages 6.6% WER.

Treat the figures as reported, not settled. The leaderboard is public, which gives its numbers more weight than a vendor PDF, and the desk has verified what Reson8 reported, not an independent reproduction. The nine languages in the average are the nine the track measures, out of more than 40 the model supports. A buyer whose audio is mostly Dutch call-center recordings or Hindi lectures should test those, not the average.

The phone-call number is the one to test

Reson8's second result uses harder audio than a leaderboard's: phone calls, compressed and noisy, in many accents. On AppTek's benchmark, 302 hours of helpline calls across 14 English accents, scored with AppTek's own script, Rumble-2-Flash reached 7.86% WER. Two models in the company's comparison did better: Qwen3-ASR 1.7B, from the same architecture family, at 7.4%, and NVIDIA's Nemotron 3.5 Transcribe, a 30B model, at 7.78%. (Reson8's announcement calls the 1.7B "the only model in the comparison that does better"; its own figures put Nemotron ahead of Flash as well.)

That ordering is the honest shape of the claim. Flash is close to a model 60 times its size on noisy, accented, compressed audio, which is the practical point. It is also behind a 1.7B model from its own family. Distillation produced a 0.5B model that competes in this range; the announcement publishes no score for the Rumble-2 teacher, so how much of the teacher's accuracy the student kept is not a number anyone outside Reson8 can check.

What the open release leaves out

Read the open-weights announcement alongside the product page, because the two describe different products. Flash transcribes recorded audio. It does not separate speakers, it does not transcribe live, and it does not accept a custom vocabulary. Reson8's page assigns all three to Rumble-2, the paid flagship, where custom vocabulary runs to 50,000 terms. The open model is the fast, cheap, multilingual front door. The features that make transcription usable in a meeting product, a call platform or a medical workflow are the ones being sold.

Apache 2.0 makes the download free, including for commercial use under the license's terms; running it still costs whatever compute the team already pays for. A team can download Flash today, run it in a standard Transformers or vLLM stack, and owe Reson8 nothing. What it cannot do with Flash is ship a diarized, vocabulary-aware transcript. Reson8's API lists transcription from €0.01 per hour of audio as a starting price. The decision that matters is whether a 0.5B model's error rate holds up on your own recordings, which neither a leaderboard nor a benchmark the vendor selected can answer for you.

Sources

  1. [1] Reson8, "Rumble-2-Flash: an open 0.5B speech model for 40+ languages," Hugging Face, November 18, 2026 — the company's own announcement: license, architecture, distillation method, leaderboard and call-center figures, speed claim, read first-hand.Read source
  2. [2] SiliconANGLE, "Reson8 releases Rumble-2-Flash speech model as a free open-weights download," November 18, 2026 — open-weights release timing, the November 3 API debut, custom vocabulary exclusivity.Read source
  3. [3] Reson8, "Meet Rumble-2-Flash," product page, updated November 10, 2026 — API starting price of €0.01 per hour of audio, recorded-audio scope, features reserved for Rumble-2.Read source