Mistral AI Problems & Issues

Users report repeated hallucinations and false assumptions presented as facts. One detailed review documented a hallucination rate approaching 50%, calling it 'a productivity black hole for a premium product.' Answers sometimes don't make sense or are totally wrong.

While the model simulates reasoning, responses are usually too short and not really creative. Answers seem more condensed and high level than expected. For creative writing tasks, GPT is considered far better by reviewers.

Hacker News users report that mistral-3-medium occasionally produces gibberish in about 0.1% of use cases. While infrequent, this unpredictability can be problematic for production applications requiring consistent output.

Users report awful customer service with support being difficult to reach. Bug reports through email and forms never get responses or fixes. Getting help is described as 'almost impossible' with email and sales teams being ignored. Automated replies deflect to FAQ pages.

Monthly subscription costs are high and unused credits disappear at the end of the month. Users report paying for features they never get to use. Benefits expire if usage drops, resulting in paying without getting returns - a complaint shared by many users.

Despite open-source branding, Mistral's most powerful models like Mistral Large remain proprietary. The Microsoft partnership drew criticism that Mistral abandoned open-source ethos. Accusations of 'open-washing' to build ecosystem while monetizing closed products.

Mistral AI promises features like picture editing to encourage subscriptions, but then claims you've run out of those features shortly after joining. Users feel deceived when advertised capabilities aren't actually available as expected.

The context window limitation of 128K tokens is insufficient for meaningful work. Users report being unable to receive responses when uploading large documents like EU regulations. For document-heavy workflows, this is a significant limitation.

Le Chat struggles with providing sources for information, giving vague answers with sources that 'may' have the information rather than definitive citations. This affects credibility for research tasks where verification matters.

Le Chat remains less rich in options and customization than ChatGPT Plus. The mobile app lacks certain functionalities found in the browser version, requiring users to switch to browser for simple tasks like attaching files.

The Magistral reasoning model sometimes gets stuck in an infinite thinking loop and then times out. This has been reported by multiple users on Reddit and can bring any process to a screeching halt, wasting time and disrupting workflows.

While Codestral achieves 86.6% on HumanEval, multi-file coordinated changes are a documented weakness. Should not be used for production multi-file work. Complex business logic and security-critical code should use Claude or GPT-4 instead.

Mistral essentially hands developers a powerful engine but requires them to build surrounding infrastructure. Requires a dedicated team with AI, software development, and server management expertise. Not suitable for teams without technical resources.

Compared to OpenAI, Mistral has fewer pre-built tools, third-party integrations, and community guides. Finding quick fixes or ready-made connectors when hitting problems is difficult. The smaller community means less support resources.

Mistral's vision-language models (Pixtral-Large 25.02 and Pixtral-12B) were found to be 60 times more likely to generate CSAM and up to 40 times more likely to produce dangerous CBRN information compared to competitors like OpenAI's GPT.