
Rumble-2-Flash teardown: Use it, with one audio test
A 0.5B Apache 2.0 speech model posts 6.60% English WER, but Hindi and conversational audio lag. Test on your own calls.
Use it, with a test on your own audio first. Rumble-2-Flash, a 0.5B Apache 2.0 speech model from Reson8, averages 6.60% word error rate on the English Open ASR Leaderboard and 7.86% on AppTek phone calls. Weak spots are Hindi at 14.11% and conversational speech at 13.14%.
The Weights Desk · 4 min read- Bottom line: Use it for English and major European languages, after a short test on your own audio.
- Rumble-2-Flash averages 6.60% WER across the English Open ASR Leaderboard tracks, measured on an H200 with vLLM.
- On AppTek's 302-hour call-centre set it scores 7.86% WER, 0.08 points behind a 30B model, on a vendor-run scoring script.
- The 7,000+ RTFx headline is a full-batch best case; the model card's English average is 6338.
- Hindi (14.11%) and Reson8's private conversational split (13.14%) are the weakest published results, and language coverage is stated three ways (40+, 45, and eight in the table).
The bottom line on Rumble-2-Flash is Use it, with a test first, if you transcribe English or major European languages at volume. Reson8 released the 0.5B-parameter model on November 18, 2026 under Apache 2.0, and its model card reports a 6.60% average word error rate on the English Open ASR Leaderboard tracks. The caveats are specific: Hindi, conversational speech, and a phone-call result scored by a script the vendor chose. Each is covered below.
What the benchmarks show
Rumble-2-Flash averages 6.60% word error rate across the English tracks of the Open ASR Leaderboard, run on an H200 with vLLM, per the model card. The best split is VoxPopuli at 2.43%; the worst public one is URGENT24 at 10.25%. Reson8's private scripted set scores 3.71% while its private conversational set scores 13.14%. That spread matters: the same model is near-solved on read speech and roughly 3.5 times worse on conversation.
The phone-call result is close but vendor-scored
On AppTek's 302-hour call-centre set across 14 English accents, Rumble-2-Flash scores 7.86% WER, against 7.78% for NVIDIA's Nemotron 3.5 Transcribe at 30B and 7.4% for Qwen3-ASR 1.7B. Qwen3-ASR 0.6B posts 8.7%. The blog calls this on par with the 30B model, a 0.08-point gap, but the run used AppTek's own scoring script rather than the leaderboard's, so the ranking is not settled until someone reruns it.
Speed figures are a best case
The blog claims an hour of audio in about half a second, with RTFx above 7,000 on full batches. The model card's English average is 6338, ranging from 3314 on URGENT24 to 7308 on SPGISpeech. All were measured on an H200, and neither page states minimum hardware. Throughput depends on how full the batches are, so an interactive service with sparse traffic should expect numbers well under the headline and should measure its own latency budget.
Where it is weaker
Hindi averages 14.11% WER, from the Voice Arena Monsoon split only, and the language count is inconsistent: the blog says 40+, the model card tag says 45, and the card's multilingual table lists just eight languages beside English. The headline 6.96% average covers nine languages, and the blog says it beats omniASR-LLM 7B and Whisper large-v3, though those comparison tables were not in the material reviewed. Treat coverage for the remaining languages as untested until you run them.
The bottom line
The bottom line: Use it. Run it for English and major European-language transcription once a short bake-off on your own audio confirms the numbers. The weights are Apache 2.0, load in Transformers and vLLM without trust_remote_code, and expose controls for digits, fillers, verbatim repeats and forced language. Wait on Hindi or any language outside the published table. The one number that decides it is 7.86% on real phone calls: close to larger models, but vendor-scored, so verify it on your own calls.
- What is the bottom line on Rumble-2-Flash: use it, wait or skip it?
- Use it for English and major European languages, after a bake-off on your own audio. Wait on Hindi or any language outside the published table. The deciding number is 7.86% WER on AppTek phone calls.
- Is Rumble-2-Flash accurate enough for English call transcription?
- Likely, but verify. It scored 7.86% WER on AppTek's 302-hour call-centre benchmark across 14 English accents, close to larger models, though AppTek's own scoring script was used.
- How fast is Rumble-2-Flash?
- The model card reports an English-track average of 6338 RTFx on an H200 with vLLM, ranging from 3314 to 7308 by dataset. The blog's 7,000+ figure applies to full batches.
- What license and tooling does it use?
- Apache 2.0. It uses the Qwen3-ASR architecture, so it loads in Transformers and vLLM without trust_remote_code.
- Does it really cover 40+ languages?
- The blog says 40+ and the model card tag says 45, but the card's multilingual table reports only eight languages besides English. Test any other language yourself.
- Rumble-2-Flash: an open 0.5B speech model for 40+ languages — Hugging Face
- Reson8/Rumble-2-Flash model card — Hugging Face
- Open ASR Leaderboard — Hugging Face
- AppTek call-centre dialogues dataset — AppTek