How to configure OpenNotebook *only* with mistral models?

[DenoBeno]

I happen to have an API key for mistral, and I would like to test OpenNotebook only with mistral models for now. Basic functionality works, but I couldn't figure out how to get TTL working. While I do see some mistral-ttl models in a list of available models, I could not use any of them in open notebook configuration.

More generally: which mistral models are recommended (or at least known to work) for which of the functions?

[VectorPeak]

For a Mistral-only setup, use one Mistral AI credential and register models by modality. Current Open Notebook supports Mistral for language, embedding, speech_to_text, and text_to_speech; Voxtral audio support landed in v1.9.0, so upgrade to at least v1.9.0 / preferably current if your install still treats Mistral as LLM+embedding only. (github.com)

The UI gotcha: Discover Models returns model names, but the Model type dropdown applies one type to everything you select in that dialog. Register Mistral models in separate passes: language models as language, embeddings as embedding, STT as speech_to_text, and TTS as text_to_speech. If a Voxtral TTS model was already registered as language, delete that model entry and re-register it as text_to_speech. (raw.githubusercontent.com)

A reasonable Mistral-only starting set is:

Chat / transformations / tools: mistral-small-2603
Higher-quality / large context: mistral-medium-3-5
Embeddings: mistral-embed
Speech-to-text: voxtral-mini-2602
Text-to-speech: voxtral-mini-tts-2603

mistral-small-2603, mistral-medium-3-5, voxtral-mini-2602, and voxtral-mini-tts-2603 are the current Mistral-side model IDs I’d try first for those slots. (docs.mistral.ai)

After registering, go to Default Model Assignments and set each default explicitly. The connection test using mistral-small-latest only verifies the credential path; it doesn’t mean that’s the only usable Mistral model.

---

If my answer solved your problem, you can click answered the question. I'm really here to help, and along the way I'm also collecting Galaxy Brain badges haha 😆

[DenoBeno]

Oh, thanks, this sounds very informative - will try. https://www.open-notebook.ai/features/model-providers.html claims mistral isn't supported for TTS and STT, while https://github.com/lfnovo/open-notebook/blob/main/README.md claims it is, so it's a bit confusing. Maybe someone can fix that discrepancy once I've tested this.

Also tried vox models already, but I didn't pay attention to what's registered as language model, need to try it again.

[DenoBeno]

Unfortunately, it didn't work. :-(

I checked and the installed version of open notebook is 1.10.0. System does not allow me to use voxtral for tts, and i'm getting error messages when I try the model test for voxtral-mini-tts-2603, for voxtral-mini-2602 and also for mistral-embed.

The last one is really confusing, since one would think that nothing can work without a model producing embeddings but I'm still somehow able to "chat with the documents". First I thought that there may be some limitation on my mistral API key, but this does not seem to be the case. Any ideas?

[DenoBeno]

In the end, the problem was on my side, not in the code. What I didn'T realise is that one has to specifically tell the system which model is used for what purpose. That is, one has to explicitly state what each of the configured models can be used for and only then default model assignments can be configured. This setup works:

That is, I have tested the models one by one and all of them return "model test passed"

[VectorPeak]

@DenoBeno sorry,dude.....my bad.....

[DenoBeno]

yeah, seems that

open-notebook/CHANGELOG.md

Line 79 in 8889087

- **Mistral Voxtral** speech-to-text (`voxtral-*-latest`) and text-to-speech (`voxtral-mini-tts`), reusing the existing Mistral credential (#827)

log entry claiming the voxtral support was added in 1.9 version was a bit overly optimistic. :-(

[DenoBeno]

in the end, the error was on my side - your answer was correct.

[VectorPeak]

@DenoBeno :-(

[lfnovo]

For a current Mistral-only setup, one Mistral credential can provide all four modalities supported by Open Notebook: language, embedding, speech_to_text, and text_to_speech.

The model classifier currently recognizes:

• embeddings: mistral-embed;
• speech-to-text: voxtral-mini-latest and voxtral-small-latest;
• text-to-speech: voxtral-mini-tts;
• language: compatible Mistral chat/language models returned by discovery.

Use Discover Models, register models under the appropriate modality, then set each required default model explicitly. I assume “TTL” in the question means TTS.

Open Notebook does not currently maintain a list of recommended Mistral language-model IDs because the provider catalog changes independently of the project. Use the current Mistral catalog and choose based on quality, context, latency, and cost. The earlier community response correctly identified the supported modalities, but its dated model recommendations should not be treated as a maintained compatibility list.

[DenoBeno]

Thanks. Right now, i am able to set-up the following models for embedding, TTS and STT. Note that I also have access to some self-hosted models, so I decided to use these for default chat, embedding and tools models. As you can see, all of these passed their respective tests.