AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineA half-billion-parameter transcriber that holds its own on noisy phone calls
· Updated · Signal Desk newsroom
Analysis of reporting by Hugging Face. How we work

Reson8’s Apache-licensed Rumble-2-Flash claims accuracy against models many times its size and extreme throughput, all on standard Transformers and vLLM. The call-center result is the one that matters most, and it is self-reported.
Key takeaways
- Rumble-2-Flash is a 0.5B-parameter speech recognition model for 40+ languages, released under Apache 2.0 with automatic language detection.
- It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM without custom code or trust_remote_code.
- Reson8 reports a 6.96% average error rate on the Open ASR Leaderboard multilingual track and 6.6% WER in English.
- On a 302-hour call-center benchmark it scores 7.86% WER, close to NVIDIA’s 30B Nemotron 3.5 Transcribe at 7.78%.
- The model was distilled from Reson8’s larger Rumble-2 using on-policy distillation across mismatched tokenizers.
Small model, big claims
Speech recognition has become one of the most useful and least glamorous corners of AI, and Reson8’s release is a good example of where it is heading. Rumble-2-Flash has 0.5 billion parameters, covers more than 40 languages with automatic detection, and comes under the permissive Apache 2.0 license. Reson8 says it competes with models three to 60 times its size and can process an hour of audio in about half a second when batches stay full, which it expresses as over 7,000 RTFx.
The practical decision that stands out is boring in the best way. By adopting the Qwen3-ASR architecture, the model works on day one in Hugging Face Transformers and vLLM with no forked runtime. For teams that have been burned by models requiring custom code, that alone is a reason to try it. Output controls, digits versus words, fillers, verbatim repeats and forced language, address real formatting headaches in transcription systems.
The phone-call test
Clean benchmark audio flatters every model. The figure we care about most is Reson8’s result on AppTek’s call-center benchmark: 302 hours of helpline calls across 14 English accents, scored with AppTek’s own script. Rumble-2-Flash reaches 7.86% word error rate, essentially level with NVIDIA’s 30-billion-parameter Nemotron 3.5 Transcribe at 7.78%, ahead of Qwen3-ASR 0.6B at 8.7%, and behind only Qwen3-ASR 1.7B at 7.4% in that comparison.
On the Open ASR Leaderboard’s multilingual track, Reson8 reports a 6.96% average across nine languages including Chinese, Hindi and English, better than Meta’s omniASR-LLM 7B and Whisper large-v3. Leaderboards are public and reproducible, which helps. But this is a vendor’s release post, the comparison set is chosen by Reson8, and the multilingual average covers nine languages, not all 40-plus. Performance on lower-resource languages is not shown.
How it was made small
The training story is worth a look for anyone building compact models. Reson8 used on-policy distillation from its larger Rumble-2: the student transcribes audio itself and the teacher scores each token, so the student is corrected on its own mistakes rather than imitating teacher transcripts. Because the two models use different tokenizers, Reson8 used a method called GOLD to align tokens before comparing them.
That approach fits a wider trend of open releases aimed at running well on modest hardware, much like the community efforts to make open-weight models easy to run. The trade-off with cheap, accurate transcription is the usual one. It makes accessibility tools, captioning and multilingual support cheaper, and it also makes large-scale recording and analysis of calls cheaper. The license does not distinguish between the two uses.
Who it affects
- Developers
- A small, permissively licensed model that drops into Transformers or vLLM lowers the cost of adding speech input.
- Contact centers
- Near-30B accuracy on noisy calls at 0.5B size could cut transcription compute sharply, if results hold on their own audio.
- Multilingual users
- Broad language coverage helps, though published accuracy covers only nine languages.
- Callers
- Cheaper transcription makes routine recording and analysis of conversations easier, raising consent questions.
What to watch
- Independent results on languages beyond the nine reported.
- How the model handles long-form audio and overlapping speakers.
- Adoption in open-source voice agent stacks.
FAQ
What license is Rumble-2-Flash under?
Apache 2.0, with weights on the Hugging Face Hub and an API option from Reson8.
How accurate is it?
Reson8 reports 6.96% average error on the Open ASR multilingual track, 6.6% WER in English, and 7.86% WER on a 302-hour call-center benchmark.
Do I need special software?
No. It uses the Qwen3-ASR architecture and runs natively in Transformers and vLLM.
More on AI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Oct 28, 2026
RTX Spark laptops get open speech models and a dial for background agentsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that react