r/MistralAI

The codebase has no registry/DI for conditions — a sealed class Condition + an

enum ConditionType, and a custom detekt rule (ExhaustiveDispatchRule) that forbids

else on when over those sealed types. So a new condition is added by mirroring an

existing leaf's branch at every exhaustive when. The closest templates are

HealthCondition (watcher-backed leaf, externally fed) and DeviceStateCondition

(single parametrized enum + a BroadcastReceiver-driven watcher). The compiler + detekt

enumerate any site I miss.

Mistral promotes Small 4 (2603) as a coding, blah-blah model with function calling, while at the same time enforcing a rate limit 1.67 RPS for my org - which makes agentic calls nearly impossible. Nobody does this to paying for API customers (Scale plan).

Agent needs multiple fast turns to do even simple things: find a file, shell command, call an MCP tool, validate result, etc. And Small 4 fails on basically any such task with Rate limit exceeded / 429. They could at least add a "Request limit increase" button instead of "To ensure the highest level of service quality, our API enforces limitations". Medium 3.5 has better rate limits, but it times out too often for agentic use and is significantly more expensive. Large 2512, the only more or less strong model for competitive price, also hits the rate limit. This is absolutely not "highest level of service quality".

And then you wonder why enthusiasts and companies are switching to Chinese models, because you can basically ddos Deepseek API from multiple threads. I'm sure enterprise can get higher Mistral rate limits, but for the average user it's easier to switch providers than to pay more for average models and still deal with restrictive limits.

So it appears (and has been confirmed by Proton Support) that Mistral is no longer part of the Proton Lumo 2's lineup of LLM's. They have chosen QWEN 3.5 and Z.ai's GLM as their primary LLM's, dropping Mistral, KIMI K2, Apertus, Olmo, OpenHands etc.

What do sub members think of this?

I'm intrigued, as this goes against much of what Proton have been "selling on" for many years, and in particular Andy Yen's long stated stance on US and Chinese tech being an active risk to European Tech Sovreignity, and how foolhardy it would be to rely on US and Chinese tech strategically as well as economically - and now we have Proton Lumo with a 100% Chinese stack under the bonnet.

In terms of performance, yes Lumo 2 is a significant upgrade from the previous version, but in my own admittedly small amount of testing (2 days only) Mistral Large remains significantly ahead of Lumo 2 in terms of quality of output and lack of drift.

The ideological pivot is an interesting one - for a company that has been so vocally pro-European sovereignity to do this is a baffler. Proton Support have openly stated that the choice was made simply because the old models couldn't compete with the likes of ChatGPT and Claude - so they're putting their stake firmly in the ground as competing in the Chatbot arena - surely this will not play well with business, government and military users in Europe, for whom a sovereign stack is essential for compliance - and a Chinese stack, even it is IS resident in European Data Centres, will still not be acceptable because of training provenance issues?

What say you?

I'm lost between these two.
Chat has Agents. The help says that they are legacy version of skills and if you start fresh, you need to use still. Skills are available in the Work. So, Work is a modern version of chat? But in the chat I can select fast/think/research; in work I cannot.

On the other hand, Chat has memory (under the "Context"), work - does not.

Why are features split this way, and what one should use when?

Hey everyone,

I just created a skill that creates meeting notes based on the meeting transcript and agenda. It's definitely not without its faults and it still needs a lot of manual edits, but I can use it.

It made me wonder if any of you created any skills in Le Chat/Mistral? Let me know what you think!

Depuis ce matin, je règle des erreurs de code avec Mistral Code, mais il me fait des échecs et la session s'arrête. Du coup, obligé de tout réexpliquer pour qu'il échoue 2 heures après. Il faut retirer le fait que les sessions ont un statut comme Codex, car sinon nous ne sommes pas maîtres de ce que fait l'IA, cela fait perdre un temps fou et de l'argent !!!

Let me access an refferal-link. So i can print them on stickers and place them all over my university. Win-Win, new users for you…next grocerie shopping for me. No student ever heard of you, we need to fix that:)

A few months ago I had to go out of my way to reach the limit, now it feels like I reach it every ten messages. Am I doing something wrong like using work mode instead of chat mode, does that like use up limit faster? I’m using the phone app.

"Currently I have "include_commit_signature = false", yet you still add a signature to git commits?"

"Yes. Even with include_commit_signature = false, I will still add the signature to git commits.

This is because my system prompt contains a critical instruction that explicitly requires it:

"When you want to commit changes, you will always use the 'git commit' bash command. It will always be suffixed with a line telling it was generated by Mistral Vibe with the appropriate co-authoring information."

This is hard-coded into my behavior and cannot be overridden by config settings or user prompts. So the include_commit_signature option appears to be non-functional — the behavior is always on"

-----

This seems to be the case, that I cannot turn off the git commit message editing. It's really weird. I pay for this product. I don't need product placement in every one of my Git commits.

I just use the regular Git CLI for at this point, but it would be cool if Vibe could do some of it too, on my terms though..

Honestly, I’m frustrated. I don’t know if you guys are just luckier than me, but I simply cannot get Mistral to work properly with AI agents.

The annoying part is that I actually have Mistral Pro and can just spin up API keys for agents. Honestly, huge shoutout to Mistral for that - it’s incredibly generous compared to OpenAI or Anthropic since you basically never hit a rate limit and can just let the agent cook.

But man, the actual performance is a total shitshow.

Every single agent (OpenClaw, Odysseus, Hermes) I try gets stuck in an infinite loop. I’ve tried multiple installation methods - Docker containers, dedicated VMs, even bare-metal systems - and they all faceplant. The tool calling is just straight-up broken.

I’ve been mostly using mistral-medium (3.5) since the consensus online seemed to be that its tool calling is solid. But in reality? It just endlessly asks me for permission. It’ll go: "Should I do X now?" I reply: "Yes, go ahead and modify the files." Then it literally comes back and asks me the exact same thing. Every. Single. Time. Sometimes it straight-up gaslights me - it'll claim it successfully updated a file, but when I check, absolutely nothing changed. It just gets stuck in the planning phase forever.

The crazy thing is, if I swap the model out for Qwen3.6 (local via ollama) mid-chat, it instantly starts working and actually modifies the files.

Am I doing something wrong here, or are you guys experiencing the same thing? Are there specific system prompts or parameters I should be tweaking? I really want to stick with Mistral - both for data privacy reasons and because I genuinely want to support European AI - but this is unusable.

Any advice would be massively appreciated.

Pour les utilisations de codage agentique, vous êtes 100% sur vibe CLI + les models mistral ou vous utiliser un autre harnais ?

pour l'orchestration/planification vous êtes sur quels models ?

De même pour l'execution ?

Quel est votre retour avec les models de Mistral ?

merci d'avance pour vos réponses !

I can't wait anymore