Voxtral – Frontier open source speech understanding models | Hacker News Reader

Voxtral – Frontier open source speech understanding models

mistral.ai

156 pointsmeetpateltech26 comments

Running Voxtral-Mini-3B-2507 on GPU requires ~9.5 GB of GPU RAM in bf16 or fp16.

Running Voxtral-Small-24B-2507 on GPU requires ~55 GB of GPU RAM in bf16 or fp16.

There is also a Voxtral Small 24B small model available to be downloaded: https://huggingface.co/mistralai/Voxtral-Small-24B-2507

24B is crazy expensive for speech transcription. Conspicuously no comparison with Parakeet, a 600M param model thats currently dominating leaderboards (but only for English)

But it also includes world knowledge, can do tool calls, etc. It’s an omnimodel

Only the mini is meant for pure transcription. And with the tests I just did on their API, comparing to Whisper large, they are around three times faster, more accurate and cheaper.

24B is, as sibling comment says, an omni model, it can also do function calling.

Im pretty excited to play around with this. I’ve worked with whisper quite a bit, it’s awesome to have another model in the same class and from Mistral, who tend to be very open. I’m sure unsloth is already working on some GGUF quants - will probably spin it up tomorrow and try it on some audio.

Does it support realtime transcription? What is the ~latency?

Unlikely. The small model is much larger than whisper (which is already hard to use for realtime)

Won't comment on the 24B model as I see no use for it personally, but regarding purely ASR tasks, I honestly can't see voxtral taking off. For personal usage, I've been running a quant of whisper tiny(for english), as well as whisper small(for spanish, as is my native language), and have never experienced major latency when using for globally available voice commands. Considering my machine runs an Ivy Bridge processor, using CPU inference, the pricing seems unreasonable.

In demo they mention polish prononcuation is pretty bad, spoken as if second language of english-native speaker. I wonder if it's the same for other languages. On the other hand whispering-english is hillariously good, especially different emotions.

It is insane how good the "French man speaking English" demo is. It captures a lot of subtleties

That’s an actual French man speaking English

My Whisper v3 Large Turbo is $0.001/min, so their price comparison is not exactly perfect.

How did you achieve that? I was looking into it and $0.006/min is quoted everywhere.

Harvesting idle compute. https://borgcloud.org/speech-to-text

Do you support speaker recognition?

No. I found models doing that unreliable when there are many speakers.

This is your service?

Yes

They claim to undercut competitors of similar quality by half for both models, yet they released both as Apache 2.0 instead of following smaller - open, larger - closed strategy used for their last releases. What's different here?

They didn't release voxtral large so your question doesn't really make sense

It's about what their top offering is at the moment, not having Large in name. Mistral Medium 3 is notably not Mistral Large 3, but it was released as API-only.

They're working on a bunch of features so maybe those will be closed. I guess they're feeling generous on the base model.

Probably not looking to directly compete in transcription space

weights:https://huggingface.co/mistralai/Voxtral-Mini-3B-2507 and https://huggingface.co/mistralai/Voxtral-Small-24B-2507

Running Voxtral-Mini-3B-2507 on GPU requires ~9.5 GB of GPU RAM in bf16 or fp16.

Running Voxtral-Small-24B-2507 on GPU requires ~55 GB of GPU RAM in bf16 or fp16.

Reply on news.ycombinator.com