---
type: "post"
id: "reson8-says-rumble-2-flash-is-now-available-7911c4"
slug: "reson8-says-rumble-2-flash-is-now-available-7911c4"
title: "Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its"
description: "Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its API, starting at €0.01 per audio hour."
retrieval_nugget: "Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its API, starting at €0.01 per audio hour. The 0.5-billion-parameter model targets inexpensive transcription across more than 40 languages; its published accuracy results are promising, but come from the company’s own evaluation."
published_at: "2026-11-21"
updated_at: "2026-11-21"
record_date: "2026-11-21"
date_kind: "published_at"
topics: ["multimodal","open-models","models"]
entities: ["Reson8","Rumble-2-Flash","Rumble-2","Whisper Large V3","Qwen3-ASR 1.7B","Nemotron 3.5 Transcribe","AppTek"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "The release details add pricing from €0.01 per audio hour and report about 92% word accuracy on AppTek’s 302-hour call-centre test."
prior_refs: []
source_url: "https://www.reson8.dev/blog/rumble-2-flash"
source_urls: ["https://www.reson8.dev/blog/rumble-2-flash"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:reson8-says-rumble-2-flash-is-now-available-7911c4"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/reson8-says-rumble-2-flash-is-now-available-7911c4/","markdown":"https://newruntime.com/posts/reson8-says-rumble-2-flash-is-now-available-7911c4.md","json":"https://newruntime.com/posts/reson8-says-rumble-2-flash-is-now-available-7911c4.json"}
markdown_url: "https://newruntime.com/posts/reson8-says-rumble-2-flash-is-now-available-7911c4.md"
---

# Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its

## Retrieval answer

Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its API, starting at €0.01 per audio hour. The 0.5-billion-parameter model targets inexpensive transcription across more than 40 languages; its published accuracy results are promising, but come from the company’s own evaluation.

## Summary
Reson8 says Rumble-2-Flash is now available as an Apache 2.0 open-weight speech-to-text model and through its API, starting at €0.01 per audio hour. The 0.5-billion-parameter model targets inexpensive transcription across more than 40 languages; its published accuracy results are promising, but come from the company’s own evaluation.

## What happened
Reson8 announced Rumble-2-Flash on November 3, 2026, describing it as its smallest, fastest and most affordable speech-to-text model. The company says the model has 0.5 billion parameters, supports more than 40 languages, and is available immediately through its API. Its weights are open under the Apache 2.0 license. API pricing starts at €0.01 per hour of audio, with no contract; at that starting rate, 1,000 hours would cost €10. Reson8 says an hour of audio can be transcribed in seconds, but does not give a single latency figure or a hardware configuration for that claim.

The release is positioned for voice notes, lectures, calls, interviews and podcasts. Reson8 says Rumble-2-Flash is designed to stay small and inexpensive while retaining accuracy on noisy and complex audio. Its evaluation includes a 302-hour AppTek call-centre test spanning 14 English accents. Using AppTek’s scoring script, Reson8 reports that the model recognized about 92% of the words. The company says this result is comparable to NVIDIA’s Nemotron 3.5 Transcribe, which it describes as 60 times larger, and better than several other named models in the comparison. It also reports that Alibaba’s Qwen3-ASR 1.7B did better. Reson8 cautions that some comparison results came from other publishers and different setups.

On the Hugging Face Open ASR multilingual test sets, the company reports a 7.09% average word error rate across German, French, Italian, Spanish, Portuguese, Hindi and Dutch. For English datasets on the leaderboard, it reports about 5% average word error rate, with results ranging from 2% on clean LibriSpeech test-clean and VoxPopuli audio to 9% on AMI meetings with overlapping speakers. On a separate comparison covering ten lower-resource languages and 17 dataset-language groups, Reson8 reports a 17.6% average word error rate for Rumble-2-Flash, against 39.8% for OpenAI’s Whisper Large V3. The company cites results including 13% for Maltese and Basque, versus 72% and 42% respectively for Whisper Large V3.

Flash does not include speaker diarization or live-conversation transcription. Reson8 says those features are planned for its larger Rumble-2 model, which it describes as its most accurate model so far and says will add speaker diarization, end-of-turn detection, automatic language detection and custom models. The announcement advises teams evaluating speech recognition to test on their own audio, and discusses how transcription benchmarks can count differences in formatting or mistakes in human transcripts as model errors.

## What the archive adds
New Runtime covered [Sarvam’s Bulbul V4 Flash release](https://newruntime.com/posts/sarvam-launched-bulbul-v4-flash-a-distilled-tts-dea6f0/) on 2026-10-15, a smaller speech-generation model presented around latency and cost constraints. Rumble-2-Flash is a new event in adjacent speech-model coverage, not a restatement: it addresses transcription rather than speech generation and publishes specific recognition evaluations.

## Trend signal
This release weakens the assumption that realtime voice models can handle tool calls, interruptions and session state well enough to replace chained speech-to-text and text-to-speech stacks. Reson8 presents Flash as a transcription component and explicitly reserves live-conversation transcription for a larger model. That does not establish which architecture is better in production, but it is a reminder that lean speech components remain relevant. See the assumption in [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/).

## Why it matters
Transcription is often an enabling layer rather than a headline capability: it turns calls, voice notes and recordings into searchable or processable text. At the announced starting price, teams with large archives can assess whether it is practical to transcribe more audio, while open weights give them an option to evaluate deployment beyond the API. The model’s reported results also make a useful distinction between average benchmark performance and fit for a particular workload. A 9% word error rate on overlapping meeting audio may be acceptable for rough notes but not for decisions that depend on exact wording. Lower-resource language results could matter to products serving speakers underserved by larger models, but teams should check performance on their dialects, recording conditions and vocabulary before relying on it. The missing diarization and live-transcription features also mean this release is not a complete voice-agent stack.

## Limits and uncertainty
The benchmark figures are reported by Reson8, the model’s maker, and have not been independently verified here. The announcement notes that some comparison bars use results from other publishers and different setups, limiting direct ranking claims. Aggregate scores also conceal differences among datasets, accents, microphones and transcription conventions. Reson8 does not specify API latency under load, throughput, hardware requirements for self-hosting, or how costs vary by usage beyond its starting price. The company says live-conversation transcription and diarization belong to the larger Rumble-2, but does not give a release date or price for that model. Teams should test their own audio and establish how errors affect downstream decisions.

## Skill to build now
Build a small evaluation set from recordings your product already handles: include clean speech, noisy calls, overlapping speakers and the languages or accents that matter to your users. Compare transcripts against corrected references, and inspect consequential errors rather than relying only on an average word error rate. If you need diarization or streaming, include those requirements in the evaluation; Flash does not claim to provide them.

## Reader takeaway
Rumble-2-Flash gives teams a low-cost, open-weight option to evaluate for batch transcription, with Reson8 reporting competitive results from several tests. Treat the numbers as a starting hypothesis, not a guarantee: validate against your own audio, language mix and tolerance for errors, and do not mistake a transcription model for a complete realtime voice system.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
