Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Daniel HarrisDaniel Harris
Heat: 560
A long teal audio waveform with a single orange accent bar on a dark field

Dek: Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Why Rumble-2-Flash matters more than a leaderboard crown

Reson8 open sourced Rumble-2-Flash on 18 November, announced on the Hugging Face blog and released under the Apache 2.0 licence. It's a speech recognition model with 0.5 billion parameters, covering more than 40 languages with automatic language detection.

The size ratio is the story. Nvidia's Nemotron 3.5 Transcribe carries 30 billion parameters. On AppTek's call-centre benchmark, which runs to 302 hours of helpline audio recorded across 14 different English accents, Rumble-2-Flash scores 7.86 percent word error rate against Nemotron's 7.78.

A tenth of a point, on phone calls, from a model 60 times smaller.

What Rumble-2-Flash scores on the Open ASR Leaderboard

The Open ASR Leaderboard is the public scoreboard most speech teams check first, and it's where a small model either earns trust or gets ignored.

On the multilingual track, Rumble-2-Flash averages 6.96 percent error across nine languages: Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. The English track average is 6.6 percent word error rate.

According to Reson8, that multilingual average beats Meta's omniASR-LLM 7B, a model about 15 times larger, and comes in ahead of Whisper large-v3.

Nine languages isn't 40, and the post doesn't break out scores for the rest. Read that figure as a floor for the big languages, not a promise for every entry on the list.

How Reson8 built Rumble-2-Flash from a larger model

The engineering is the part buyers skip, and it's the reason this model is small without being useless.

Rumble-2-Flash is a distilled student of Reson8's larger Rumble-2 production model. Distillation means a small model learns by imitating a big one, which usually produces something cheap and mediocre. The difference here is the grading loop: the student transcribes the audio itself, and the teacher scores every token it produces. That's on-policy distillation.

There was a snag. Teacher and student use different tokenizers, the chunks of text a model reads and writes, so comparing their outputs directly tells you nothing. Reson8 solved it with a method called GOLD, which merges tokens until both models are graded on the same text.

Skip that detail and the result looks like magic. Keep it and the result looks like a fortnight of careful work.

Speed: 7,000 RTFx and an hour of audio in half a second

RTFx measures how much audio a model can process per second of compute, so a higher number means less waiting.

Rumble-2-Flash reaches over 7,000 RTFx on full batches, which Reson8 describes as roughly an hour of audio in about half a second. A support team sitting on 20,000 recordings from the past week can run the lot overnight on hardware it already owns, which is usually the same machine that sat idle between midnight and six.

Speed changes what you can build. Live captioning stops being a batch project.

Where Rumble-2-Flash still loses on call-centre audio

Parity has a boundary, and it's worth naming it before anyone plans a migration.

On AppTek's 302-hour set, Rumble-2-Flash lands at 7.86 percent, behind Nemotron at 7.78 and behind Qwen3-ASR 1.7B at 7.4. It does beat Qwen3-ASR 0.6B, which sits at 8.7.

So the accurate claim is that a 0.5B open speech model reached the neighbourhood of a 30B model on messy phone audio. It didn't win the category, and it didn't beat every model its own size.

Clean studio audio is a different test. The model card doesn't pretend otherwise.

What the Apache 2.0 licence changes for vendors

Apache 2.0 allows commercial use, modification and redistribution, with no per-minute fee, no seat count and no clause that stops you from shipping the model inside a paid product of your own.

No licence fee, then.

For a vendor that resells transcription, that removes a line item and a procurement conversation at the same time. It doesn't remove the operational work. Someone still has to run the GPUs, handle retries, store recordings legally and watch quality after an update.

What deploying an open speech model actually looks like

The boring compatibility list is what separates a demo from a deployment.

Rumble-2-Flash runs natively in Transformers and vLLM, with no custom code and no trust_remote_code flag. For most teams that's the difference between an afternoon and two weeks of yak shaving. The architecture follows Qwen3-ASR, so existing pipelines have somewhere to start.

Output controls matter more than they sound, and they're the reason transcripts from this model can feed a billing system without a cleanup script in between: numbers as digits or as words, fillers kept or dropped, verbatim repeats preserved, a forced language, or a parsed output when the model detects language on its own.

Call-centre vendors argue about exactly this in procurement, because downstream systems break on "two" when they expect "2".

Who should care about a 0.5B open speech model

Two groups will feel this first, and neither of them is a research lab.

The obvious one is any team paying per minute for transcription at volume. At current API rates, a few hundred thousand minutes a month is a real budget line, and a model you can host yourself turns it into a fixed cost.

The second is anyone who can't send audio off site. Hospitals, law firms, insurers and public agencies have recordings they're not allowed to upload, and a 0.5B model that runs on a single GPU changes that conversation.

On-device is the third possibility, and it's the one to be careful about. Reson8 documents sizes and throughput, not a phone benchmark, so nobody has shown this running well on a mid-range Android handset yet.

So who should test it this week? Anyone with a transcription bill above four figures a month.

Weights are on Hugging Face at Reson8/Rumble-2-Flash, and the model is also available through Reson8's API. If you're evaluating it, test on your own worst audio first, the accented, echoey, crosstalk-heavy calls that break transcription in production. That's where the tenth of a point stops mattering and the error types start to.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
Three neon rings orbiting a small glowing core on a deep blue field
Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Heat: 1,500
A large model core ringed with reasoning spokes beside a smaller fast core, with falling price bars on the right
Models

OpenAI's GPT-6.2 Astra Is 20% Cheaper and Still Not the Best at Everything

OpenAI cut the flagship's token prices and left the cheap model where it was. The leaderboards OpenAI doesn't run put the expensive one seventh.

Heat: 1,900