Mistral released new models:
https://mistral.ai/news/mistral-small-4

Small evolved into a 119B MoE model with 6B active parameters. (Small 3 used to be a dense 24 GB model instead.)

Another new model is Leanstral, which serves as a proof assistant:
https://mistral.ai/news/leanstral

---

I have been switching from Devstral Small 2 to Qwen Coder Next; but since Mistral Small 4 can be seen as a successor to Devstral Small, will be interesting to compare.

A setup that I have been using recently is to have Mistral Vibe running in a Docker container without network, mounting only a writable copy of the current working directory. 127.0.0.1:8080 is exposed, so it can still connect to a local API (llama.cpp in my case).

Since this setup reduces the blast radius, I can allow Vibe to use tools without confirmation.

Not sure if it is useful for someone else - it is tailored to my own use case and will likely not work outside of Linux - but I put the code here:

https://codeberg.org/philipp-classen/vibe-in-docker

#mistral

Nice, in-depth article on anonymous-credentials and its challenges:

https://blog.cryptographyengineering.com/2026/03/02/anonymous-credentials-an-illustrated-primer/

Anonymous credentials and the related term zero-knowledge proof are core building blocks for privacy-preserving online age verification.

As I understand, the age verification Australia currently requires to upload a legal document (e.g. a driver's license or government ID), or requires proof by doing a "video self". Both techniques would not be privacy-preserving, because in both cases you are sharing personal information with a company (e.g. Meta).

Cryptography like anonymous credentials solves that problem. At least in theory. You can prove your age, and the other side does not learn anything else about you. Still, to roll it out in real-world scenarios there are many challenges (e.g. dealing with attacks like credential theft).

Side-note: the EU proposal is also built on the same type of cryptography: https://ageverification.dev/av-doc-technical-specification/docs/architecture-and-technical-specifications/

#crypto #AnonymousCredentials #ageverification

Data sovereignty and the ability to exercise rules for the big online platforms are important for many reasons:
* Large social media platforms allow countries like Russia to spread disinformation to destabilize Western democracies and raise concerns about mental health issues.
* Data privacy challenges arise when sensitive data is shared with US or Chinese companies.
* Much critical IT infrastructure depends on US companies. (As shown by Microsoft blocking the email of the International Court of Justice, this is no longer a theoretical concern.)
* Same with financial services - booking a hotel in Europe without using US financial providers? Currently not possible. (I have only seen credit card or PayPal so far.)

Legislation like the DSA was an important step in 2022; but in the last year it feels like the will to move to alternatives like European providers or Open Source solutions increased in the society. That is promising, though there is a long way to go.

But not surprisingly, it is getting more confrontational now. I hope that ongoing pressure like that will not be successful:

https://www.reuters.com/sustainability/boards-policy-regulation/us-orders-diplomats-fight-data-sovereignty-initiatives-2026-02-25/

Good overview of the large language models published since 2017:
https://llm-timeline.com/

#llm #ai

Mistral released new coding models:

* Devstral 2 (123 B)
* Devstral Small 2 (24 B)

https://mistral.ai/news/devstral-2-vibe-cli

Plus, Mistral Vibe CLI, a CLI based coding agent:
https://github.com/mistralai/mistral-vibe

I didn't even have a chance yet to try out Mistral 3 Large. According to their benchmark, Devstral 2 is competing with the best open weight models (DeepSeek V3.2). Not quite at the level of the best closed-sources ones, but more cost-efficient:

> Devstral 2 is currently offered free via our API. After the free period, the API pricing will be $0.40/$2.00 per million tokens (input/output) for Devstral 2 and $0.10/$0.30 for Devstral Small 2.

Devstral Small 2 can be run locally on a 24 GB GPU.

#mistral #ai #devstral

Mistral.ai had a major release: https://mistral.ai/news/mistral-3

Three smaller, dense models and Mistral 3 Large, a large mixture-of-experts model (41B active and 675B total parameters). According to the first benchmarks, this puts it in the same league as the best open models (currently all Chinese).

Mistral's release includes the base model, not just the fine-tuned version as it had become more common. Plus, they used an Apache 2.0 license. That combination will likely open up fine-tunes from the community and makes it more useful for companies.

Mistral 3 Large is not a reasoning model, but they mentioned that a reasoning version should be coming soon.

I didn't have the chance to try them out myself. For specific problems like German content, I personally found Mistral 2 Large (from 2024) still doing better than the top Chinese ones (DeepSeek, Qwen, Kimi, GLM), even if they otherwise are very capable.

Mistral 3 Large has been released here:
* https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512
* https://huggingface.co/mistralai/Mistral-Large-3-675B-Base-2512

#mistral #ai #llm

I think, the "btdu" AUR package has been compromised with the last commit.

https://aur.archlinux.org/packages/btdu

Similar to this:
https://bbs.archlinux.org/viewtopic.php?id=309861

Trying to escalate, but sending to "[email protected]" was rejected.

@archlinux @anthraxx

#arch