Mistral Gave Away Its Biggest Model

Mistral put its biggest model out in the open

Mistral released the weights for Mistral Large 4 on October 29. Weights are the trained settings inside a model. When a company shares them, anyone can download the model and run it. TechCrunch reported the release, and Mistral posted the model on Hugging Face.

The model is large by any measure. It holds about 1 trillion parameters in total. Only about 50 billion of them are active for each piece of text it handles. It can also take in text and images, and it can hold a very long conversation in memory.

Mistral built this one for coding and for cyber defense. The company says it scores about 51% on a cyber test index. That test checks if a model can find flaws in code and patch them. Mistral also says the model refused about 97% of harmful cyber requests in its tests. Those are the company's own numbers, and outside checks will matter.

The open part is what changes the choice for a lot of teams. A hospital or a bank can run this model on its own servers. Its files never have to leave the building. The license is a changed form of the MIT license. People can use it and build products with it. Companies that sell it as a hosted service need a separate deal.

Mistral was open about the risk too. The company said it saw a chance the model could help with real attacks. So it is tightening its guardrails in the months ahead. A stronger version made for penetration testing goes only to vetted partners. Penetration testing is when experts try to break into a system on purpose, with permission.

Security groups are already trying it. Booz Allen, CrowdStrike, and Elastic have been testing the model. Europe's cyber agency, ENISA, is testing it as well. For a regular reader, the takeaway is simple. A top model for finding holes in software is now free to download. That helps defenders who could not afford one before. It also means the same tool sits on the open internet.

Gemini now lets you fix the file inside the chat

Google added file editing inside Gemini on October 29. You can now make a file in a chat, see it there, and fix it there. The Verge reported the update, which Google shared on its Gemini blog. It is open to all Gemini users, in every language the app supports.

Before this, Gemini could make a file and hand it to you. If a chart was wrong, you opened the file somewhere else. Or you asked Gemini to make the whole file again. Now the file opens right in the conversation. You can expand it to full screen and type your own changes.

You get three ways to edit. You can type changes yourself in the full screen view. You can highlight a part and ask Gemini a question about it. Or you can ask Gemini to rework that part for you. Google gave one clear example. A student uploads photos of handwritten notes. Gemini turns them into a PDF study guide, and the student edits it in place.

There is one limit to know. PDF files work on both the web and the phone app. Docs, Sheets, and slide decks work on the web only for now. Markdown files, text files, and code are also web only. So most real work with these files still happens at a desk.

Try this the next time Gemini makes a file for you. Ask it for a short doc or a simple sheet, then stay in the chat. Highlight one weak section and ask Gemini to tighten just that part. Then open the file full screen and fix one line yourself. You will learn fast where the tool helps and where your own eye still matters.

Agents are now building the test worlds

NVIDIA shared a set of projects on October 29 where agents build simulations. A simulation is a test world on a computer. Teams use one to try a design before they build the real thing. NVIDIA posted the examples on its developer blog.

The work split is easy to picture. A developer describes the scene in normal words. The agent gathers the parts, links the physics, and checks the scene. Omniverse, NVIDIA's simulation platform, supplies the physics and the images. The human reviews the work and asks for changes.

The examples cover real jobs. One team turned site scans into a clean model of a building. Workers can move shelves and machines in the model first. Another team built a what-if tool for a gas turbine. Others tested warehouse layouts, robot arms, and self-driving cars on a street in San Francisco.

NVIDIA shared one warehouse test in detail. The scene ran smoothly for two minutes at full HD quality. It held 80 conveyors, 40 arms, and 800 boxes. When the team doubled the scene, it slowed down. That kind of limit matters. These are company demos, not outside tests, and the post says so in its own way.

Still, the direction is worth watching. Building a good test world used to take weeks of setup work. If an agent can do most of that setup, more teams can test ideas early. A small firm can ask what breaks before it spends real money. That is where this gets useful beyond big labs.

Also today, Dream came out of stealth with about $17 million in seed funding led by Andreessen Horowitz, TechCrunch reported. The startup builds AI models for building design, trained on permit records for over 100 million square feet. Its tools are in use on 10 projects in Texas and California. NVIDIA also said its thin RTX Spark laptops gained new models and agent tools this week. Microsoft put its Copilot harness for building agents into public preview in Foundry Agent Service.

Today, a top model went open, a chat tool learned to edit files, and agents started building test worlds. Each step hands you more of the work itself. The skill that grows with it is checking the work before you trust it.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Product details change fast; check primary sources before you build on them.

Want more with the Tech Simplified crew? Subscribe on Substack.

Get the week's briefs in one email. The Briefcase packs Monday to Friday's AI Briefs into one Sunday email.

Follow me on LinkedIn for project updates, ideas for using AI, and lessons you can bring to your own work.