Overview
Rumble-2-Flash is a speech recognition model Reson8 released as open weights on November 18, 2026. Apache 2.0, 0.5 billion parameters, built on the Qwen3-ASR architecture. It runs natively in Transformers and vLLM on release day, with no custom code and no trust_remote_code.
The training method is on-policy distillation. The teacher is the production Rumble-2 model, the student transcribes the audio itself, and the teacher then produces its own next-token distribution at every position for the student to match. Corrections therefore target the student's own mistakes rather than having it imitate a model that tokenizes differently. The tokenizer mismatch is handled with GOLD, which merges tokens the two models split differently and compares shared tokens and their sorted probabilities.
Results come in two parts. On the Open ASR Leaderboard multilingual track, covering Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, the average error rate is 6.96 percent, with 6.6 percent on the English track. Meta's omniASR-LLM 7B on the same table is 15 times larger, and Whisper large-v3 is also in the comparison. Reson8's own framing is that it competes with models 3 to 60 times its size.
The second part is call centers, which says more about the model. AppTek's benchmark uses 302 hours of hotline calls across 14 English accents, scored on AppTek's own transcripts: Rumble-2-Flash lands at 7.86 percent, NVIDIA Nemotron 3.5 Transcribe at 30B gets 7.78 percent, Qwen3-ASR 0.6B gets 8.7 percent, and Qwen3-ASR 1.7B at 7.4 percent is the best row in the table. Telephony audio is close to the hardest input speech recognition faces, and at that difficulty the model sits in the same band as 30B models while being roughly two orders of magnitude smaller.
Speed is over 7,000 RTFx, which Reson8 converts to about half a second per hour of audio. That figure assumes full batches, as in the larger datasets on the Open ASR Leaderboard; throughput drops when batches are not filled, so benchmark it on your own traffic shape.
On the engineering side it is low friction. Transformers and vLLM both support it natively, and device_map set to auto handles placement. Output can be controlled: digits as numerals or words, whether to keep fillers, whether to keep verbatim repeats, and a forced language. Audio input accepts local paths, URLs, NumPy arrays or a list of them for batching. Language detection is automatic by default, and coverage runs past 40 languages.
Limits are worth stating plainly. The blog does not give memory or VRAM requirements, and it does not describe the training data composition or hours, so questions about running on a phone are not answered there. Posts on X describe a 1GB model with 45 languages and offline transcription on a handset; the blog itself says more than 40 languages and says nothing about size. The two do not line up, and this page follows the blog.
Who it is for. Teams that need transcription inside their own environment, that batch long audio, or that run call center QA and meeting notes. At 0.5B the model fits on a single consumer GPU or even a CPU, and Apache 2.0 removes the commercial licensing conversation. If you only transcribe a few recordings now and then, a hosted API is less work.
Key Features
- 0.5B parameters under Apache 2.0: A 500-million-parameter model with a permissive commercial license, deployable in a private environment or embedded in a shipping product. Weights live under the Reson8 organization on Hugging Face.
- On-policy distillation: Distilled from the production Rumble-2. The student transcribes first, the teacher supplies a per-position next-token distribution, and the student matches it. Corrections target the student's errors rather than copying teacher output.
- 6.96 percent average multilingual error: Nine languages on the Open ASR Leaderboard multilingual track average 6.96 percent, with 6.6 percent on English. Meta's omniASR-LLM 7B on the same table is 15 times larger.
- 7.86 percent on call center audio: AppTek's benchmark covers 302 hours of hotline calls across 14 English accents. NVIDIA Nemotron 3.5 Transcribe 30B scores 7.78 percent and Qwen3-ASR 1.7B scores 7.4 percent on the same table.
- Over 7,000 RTFx: Reson8 converts that to roughly half a second per hour of audio under full batches. Throughput falls as batch fill drops.
- Native frameworks and output control: Native Transformers and vLLM support with no custom runtime. Controls cover digit format, fillers, verbatim repeats and forced language, and audio can arrive as a local path, URL or NumPy array.
Use Cases
- Batch transcription and QA for call center recordings, where accented speech, overlapping voices and line noise dominate
- Local transcription of meetings and voice memos without audio leaving the internal network
- Long-audio batch pipelines served through vLLM to hold throughput
- Multilingual content archiving that needs automatic language detection before downstream routing
- Embedding transcription into your own product without a licensing negotiation, under Apache 2.0
Pros
- At 0.5B the model fits on a single GPU, which keeps the per-hour cost of long-audio batch work low
- Error rates on call center audio sit close to 30B models while the model is two orders of magnitude smaller
- On-policy distillation makes corrections target the student's own mistakes instead of copying teacher output
- Transformers and vLLM work out of the box, with no trust_remote_code and no runtime fork
- Output controls are fine-grained enough to match downstream needs, including digit format, fillers and verbatim repeats
- Apache 2.0 means no extra commercial licensing process
Pricing
Weights are published on Hugging Face under Apache 2.0, so downloading and commercial use are free and the real cost is inference hardware plus operations. At 0.5B, inference fits on a single consumer GPU or a CPU, and for batch transcription the cost is mostly electricity and machine time. Reson8 also runs a hosted endpoint at reson8.dev billed per API call, which suits teams that would rather not operate a service or only transcribe occasionally. Self-hosting usually wins on marginal cost for steady volume; a hosted API wins when volume is sparse and operations overhead matters more. Check the official page for current list prices.
Summary
Rumble-2-Flash compresses speech recognition into 0.5B parameters while holding error rates near 30B models on telephony audio. 6.96 percent average multilingual, 6.6 percent English, 7.86 percent on the call center benchmark, over 7,000 RTFx, Apache 2.0, native Transformers and vLLM. Its value is not a new leaderboard record but an order-of-magnitude drop in the cost of private deployment, long-audio batching and in-product transcription. Two caveats: memory requirements and training data composition are unpublished, and claims about phone-side offline transcription currently rest on third-party posts rather than a primary source.
Version History
- Rumble-2-Flash released as open weights (2026-11-18): Reson8 released a 0.5B speech recognition model under Apache 2.0, built on the Qwen3-ASR architecture and trained by on-policy distillation from production Rumble-2, with GOLD aligning the tokenizers. 6.96 percent average error on the Open ASR Leaderboard multilingual track, 6.6 percent on English, 7.86 percent on AppTek's call center benchmark, over 7,000 RTFx, and native Transformers and vLLM support.