Realtime AI News
Reson8 open-sources Rumble-2-Flash, a 0.5B speech model for 40+ languages
Reson8 released Rumble-2-Flash on November 18 under Apache 2.0: a 0.5B-parameter speech recognition model covering more than 40 languages that runs natively in Transformers and vLLM. It posts a 6.96% average error rate on the Open ASR Leaderboard's multilingual track, beating Meta's 15x larger omniASR-LLM 7B and matching a 30B Nemotron model on a 302-hour call-centre benchmark.

Reson8 used its Hugging Face blog on November 18 to open source Rumble-2-Flash under the Apache 2.0 license: a 0.5B-parameter speech recognition model covering more than 40 languages. The company's stated goal is to compete with models three to 60 times its size.
The most immediate change is on the deployment side. Because Rumble-2-Flash is built on the Qwen3-ASR architecture, it runs natively in Transformers and vLLM from day one, with no forked runtime and no trust_remote_code. Weights are on the Hub at Reson8/Rumble-2-Flash and the model is also offered through Reson8's API.
Functionally, the model detects language automatically and lets callers control the output format: numbers as digits or as words, whether to keep fillers, whether to preserve verbatim repeats, and an option to force a specific language.
Accuracy is the headline claim. On the multilingual track of the Open ASR Leaderboard, Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That beats Meta's omniASR-LLM 7B, a model 15 times larger, and sits ahead of Whisper large-v3; on the English track it averages 6.6% WER.
Speed matters just as much for the use cases Reson8 is targeting. On the larger datasets in the leaderboard, where batches stay full, it reaches over 7,000 RTFx, which the blog describes as an hour of audio in about half a second.
The harder test is telephony. On AppTek's call-centre benchmark — 302 hours of helpline calls across 14 English accents — Rumble-2-Flash records 7.86% WER, essentially level with NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78% and well ahead of the 0.6B Qwen3-ASR at 8.7%. Only the 1.7B Qwen3-ASR, at 7.4%, does better in that comparison.
The method behind the size reduction is on-policy distillation from Rumble-2, Reson8's larger production model. The student transcribes audio itself and the teacher grades that transcript token by token, so corrections land on the student's own mistakes rather than on teacher-written transcripts. Because the two models use different tokenizers, the team introduced GOLD, which merges tokens that are split differently so both sides are compared over the same text — removing the shared-vocabulary requirement that usually constrains on-policy distillation.
For the speech stack, this is a signal that competition is shifting from raw scale to a practical balance of size, latency and multilingual coverage. The 0.5B footprint means it can run on a single GPU or at the edge, while public weights under Apache 2.0 remove licensing friction for fine-tuning. What to watch next is whether the accuracy holds on messier, noisier real-world audio, and whether small open ASR models start pushing down the price of commercial transcription APIs.
Why it matters
The release shifts speech recognition competition from raw scale toward latency, language coverage and deployment cost, and cheap open weights could pressure commercial transcription pricing.
Nearby Updates
All11/18, 08:00
OpenAI and Groupe Figaro partner to bring Le Figaro journalism into ChatGPT
OpenAI and Groupe Figaro announced a strategic partnership that makes Le Figaro's current reporting and archive content discoverable inside ChatGPT through summaries and excerpts with links to the original stories. Both sides said the publisher keeps full editorial independence and control over its journalism.
11/18, 08:00
OpenAI details model-driven chip design: 13% area savings on Jalapeño, 14% PPA gain on AMD GPUs
OpenAI described how its models cut standard-cell area by roughly 13% in Jalapeño's warp engine and 10% in its matrix unit, and delivered up to 14% PPA improvement over heavily optimized baselines on AMD's flagship GPU IP. It also reported GPT-6.2 Astra leading evaluations on two open-source RISC-V cores and a new partnership with Synopsys to train GPT-Synopsys on EDA tools.
11/18, 08:16
Nvidia says record revenue run won't stop, even amid leadership turmoil
Nvidia said its string of record-breaking revenue is not about to end, even as the company navigates the fallout from founder and CEO Jensen Huang's departure, TechCrunch reported. The report attaches a caveat to the optimism: the projection is framed as holding for now.
11/18, 07:37
Nvidia's interim CEO says chip sales to China have stopped, TechCrunch reports
TechCrunch reported on November 17 that Nvidia's interim CEO has said the company is no longer selling chips to Chinese customers, framing the statement against a leadership upheaval. The story asks what the decision will cost but gives no revenue figure or timeline.