r/MistralAI

I am new to Mistral AI. There are 3 models, and often I am not sure what models to use for most effective response. I read help guide and there are certain examples of when to use Thinking vs bulk task with Fast vs Deep Research. I don't want to waste limits. Is there a good rule of thumb for easier understanding? I always use Thinking and rarely Fast since I get bias due to Hallucination, but at times, I thought it is wasteful as well. Just curious on how others rationalize what models you select.

Who has already used it? What are your opinions?

https://preview.redd.it/dolk6c3v82tg1.png?width=2557&format=png&auto=webp&s=49fcd2c47054e569c978ef49e5f981bf74bdc756

I am an AI sceptic. However, I don't want to be left behind, so I've been experimenting with Claude Opus 4.6 lately through my org's Copilot subscription, and it's the first model that I tried that actually seems capable for coding and performing semi-complex tasks (with the right guardrails).

But, I would never finacially support a company like Anthropic, so I got a Mistral Vibe subscription yesterday. I fully expected it to be slightly less capable, and I'm fine with babysitting the agent a bit more. Clause Opus does make you lazy, so probably a good thing to be a bit more involved.

Unfortunately, Devstral 2 seems to be basically unusable for basic operations. Here are some examples of my first interactions with this model right after setup:

| Q: What mcp servers do you have access to?
| A: I don't have access to any Minecraft (MCP) servers. My capabilities are focused on code analysis, editing, and development tasks within your local codebase.

It doesn't know what MCP means...

| Q: Can you convert the following VSCode MCP server configuration into a Mistral Vibe MCP config?
| A:

servers:
github-mcp-server:
	type: http
	url: "https://api.githubcopilot.com/mcp/"
	gallery: "https://api.mcp.github.com"
	version: "0.33.0"

context7:
	type: stdio
	command: "npx"
	args:
	- "@upstash/context7-mcp@1.0.31"
	env:
	CONTEXT7_API_KEY: "${input:CONTEXT7_API_KEY}"
	gallery: "https://api.mcp.github.com"
	version: "1.0.31"

gopls:
	type: http
	url: "http://localhost:8092"

inputs:
- id: CONTEXT7_API_KEY
	type: promptString
	description: "API key for authentication"
	password: true

It doesn't know it's own config is in TOML format, and just hallicinated something instead.

| Q: What agent instructions are provided to you by this repo?
| A: 1. Use rtk prefix for supported commands
| Q: So, when an AGENTS.md is present in the current workplace, you will ALWAYS follow these guidelines?
| A: Yes, exactly.
| Q: I have seen you use `grep` a bunch of times, without prefixing it with `rtk` first as instructed.
| A: You're absolutely right to call this out.

It doesn't follow AGENTS.md instructions, even though it claims it does.

| Q: How can I clear context in current session?
| A: Use session management /session new

This doesn't exists, it doesn't know the commands of its own CLI.

I have also tried to make it implement a bit more complex feature, which it got stuck on for 30 min. I solved it in 15 min. Clause Opus could have probably fixed it even faster.

Am I doing something wrong? How can you use a tool for coding if it cannot even guide me through its own setup correctly?

I've now setup MCP server, skills, and agents instructions, so hopefully it will start to perform slightly better.

Are there other steps I can take to improve performance of this tool? I'd really like to keep my subscription and make this work.

Anyone else's flash answers gone missing?

So are few months back there was the Mistral Small Creative experimental model that was meant to be specialized for creative purposes that looked quite promising.

Does anyone know if there is a planned open weights release for that model? Or what is the parameter of that model?

We moved from Open AI API to Mistral because users had concerns over GDPR. Now we are facing daily downtimes on their Chat completions API. Anyone else concerned about this?

Maybe someone has some good tips?

Also their incident reporting is not very useful:
https://status.mistral.ai/

Updates often happen in retrospect and are not precise at all.

Please Mistral-Team, we like your models and offerings. But please get that fixed asap or we will leave again for Anthropic. This situation is unacceptable for our software business. Please inform your clients about the current situation and what you are planning to fix this.

We went to our 1000 products Woocommerce demo blog, we searched for the cheapest coffee maker for 4 persons.

As you can see, the cheapest coffee maker in the demo shop is Series 117 for 42 USD.

This is the cheapest coffee maker in our DEMO e-commerce

So we asked Fabio AI Chatbot for a coffee maker for 4 persons cheaper than 30 USD

1- Mistral Large 3

https://preview.redd.it/9369lqwg0ysg1.png?width=962&format=png&auto=webp&s=9cbb917993e3b1be4970a3f992b8ce85439d1904

Mistral large got it right : there are no coffee maker at this price and for 4 persons, but it directed the user to the cheapest corresponding coffee maker which can indeed make coffee for 4 persons 👍👍.

2- Mistral Medium 3.1

Mistral Medium 3.1 got it nearly right : it points to a 4 persons max coffee maker but not close to requested price :

https://preview.redd.it/u1ooxlk43ysg1.png?width=927&format=png&auto=webp&s=29046d1b694e280ddd76c6f3710ae39190db44ef

Indeed, model 169 gray in our demo store is flagged for 4 persons max, but nowhere close to the requested price.

https://preview.redd.it/2y60zeuc3ysg1.png?width=1472&format=png&auto=webp&s=c31c3037fc119d4c14272759ffd729c081d118a7

3- Mistral Small 3.2

https://preview.redd.it/itda07wd3ysg1.png?width=968&format=png&auto=webp&s=3621272e64abc0f37e7adc5fe35f4fa1fd2803e0

Mistral Small 3.2 gives the same answer as Mistral Medium 3.1: it finds a coffee maker for up to 4 person, but far from the requested price.

Conclusion : in this case, Mistral Large gives the result a e-shopper would expect.

How about you, which AI model would you recommend for this kind of search on a 1000 e-commerce products ?

[ Removed by Reddit on account of violating the content policy. ]