Lilith.
⌕
Editorial illustration: Rumble-2-Flash packs transcription for 45 languages into 0.5B parameters
Lilith illustration · editorial remix

Reson8 has released the Rumble-2-Flash weights under the Apache 2.0 license. The model has 0.5 billion parameters, its metadata lists 45 languages and the Qwen3-ASR architecture lets it run directly in Transformers and vLLM without trust_remote_code.

A half-billion parameters target transcription without a custom runtime

Rumble-2-Flash was distilled from the larger production model Rumble-2. The student transcribes the audio itself and the teacher scores its output token by token. The GOLD method allows the two models to be compared despite their different tokenizers.

The model can detect language automatically and control output details such as digits, numbers written as words, fillers and repeated words. Reson8 offers a paid API alongside the open weights. The model card does not state minimum GPU memory or the maximum recording length.

The small footprint changes the economics of private transcription

Half a billion parameters with standard Transformers and vLLM support lowers the barrier for teams that do not want to send calls or meetings to another company's cloud. Cost is only part of the appeal. Local operation gives the owner control over audio retention, latency and model updates.

Rumble-2-Flash averages 6.60% WER on the English Open ASR Leaderboard. On 302 hours of calls covering 14 English accents, Reson8 reports 7.86%, compared with 7.78% for NVIDIA Nemotron 3.5 Transcribe at 30 billion parameters. That is relevant to call centres, but it remains one specific test set.

The 7,000 RTFx headline requires full batches and an H200

The claim that an hour of audio can be processed in roughly half a second comes from throughput above 7,000 RTFx with full batches. The model card says the leaderboard run used an H200 with vLLM. Sparse traffic on a smaller GPU should not be expected to reproduce that figure automatically.

Accuracy also varies sharply by dataset. English split results range from 2.43% WER on VoxPopuli to 13.14% on the private conversational set. Chinese is measured with CER, so the combined multilingual average is not a simple comparison of one identical metric.

Real recordings will matter more than the leaderboard average

Deployment should be decided with actual calls, microphones, names and specialist vocabulary. Teams should measure small-batch latency, memory use and errors on numbers or proper names alongside WER.

The stability of Transformers and vLLM support will matter too. If the model runs without forks on available hardware and retains quality in a local audit, it could replace much larger transcription models. A card listing 45 languages does not guarantee that result by itself.

Lilith's verdict

Rumble-2-Flash fits in a smaller box, but a meeting-room microphone does not care about leaderboard tables. Real calls will quickly reveal whether this is a compact engine or merely a lightweight bonnet.