LiteRouter Academia 

Three weeks is a short time in model development, yet that was all Mistral needed to move Large 4 from 28% to 41% on a coding test that measures how well a model works inside a terminal. The extra training was guided by feedback from governments and enterprises, and it also changed how the model spends its effort: users can now pick low, medium, or high reasoning, and the high setting got 50% more efficient along the way.

The efficiency story goes further than speed. Mistral reports that the model reaches results comparable to GLM-5.3 while using 47% fewer tokens per task, a gap that grows quickly when coding agents run for hours. The article explains what tokens are, why they drive cost, and how the cybersecurity test behind the 79% score works, including the bugs it asks the model to find and patch.

There is also a debate attached to the release. Mistral's vice president of science compares AI to electricity and argues that nobody should depend on a supplier who can cut off access, while critics point out that published weights can never be recalled. The full piece lays out both positions, along with what it takes to run a trillion-parameter model on real hardware.

Before October 19, an anonymous model called Space Bunny had been handling more than 200 trillion tokens, and nobody could say who made it. MiniMax has now confirmed that Space Bunny was an early build of M3.1 Flash. A preview had already appeared inside the company's coding app at the end of September, with no price and no benchmarks attached, so launch day brought the numbers.

Those numbers tell a mixed story. The price is flat at $0.15 per million input tokens and $0.60 per million output tokens, with a context window of one million tokens and speeds above 200 tokens per second, according to the company. On BankerToolBench, the model ties Opus 5, and on a coding test called DeepSWE it edges GPT-5.6 Sol, yet on Terminal Bench 4.0 it trails Opus 5 by more than thirteen points. Every score comes from MiniMax, and MarkTechPost found that one rival's results on the same Automation-Bench look quite different depending on who ran the test.

Two design choices deserve attention. Reasoning cannot be switched off, so even a simple request carries a minimum of tokens and waiting time, and the model offers five effort levels, with Max recommended. The open weights, meanwhile, carry a promise without a date or a license on the company's page, although one outlet lists October 23. If that day arrives, independent testers will finally be able to check the claims.

Mistral AI announced Mistral Large 4 on October 6, a model with one trillion parameters in total. Only 49 billion of them work on any single request, which is how a model this large keeps the cost of each answer under control. The company trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its data centers in France, and the same cluster now serves the public preview. The article walks through how that design works using a simple hospital comparison, so the idea of experts switching on and off becomes easy to picture.

The company reports 62% on DeepSWE v1.1, a coding benchmark, a little ahead of several well-known open-weight models, and it pitches the model heavily for cybersecurity work. Mistral itself calls the benchmark figures preliminary, and the cyber claims arrived without published numbers. The article keeps those two things apart, so you can see what has been confirmed by several outlets and what still rests on the company's word.

The weights are scheduled for October 27, and that is the date when researchers and companies can start testing the model themselves. We also explain what open-weight means in practice and how the €3 billion Series D round connects to the cluster behind this release. If you follow GPUs and data centers, the training setup is probably the most interesting part of the story.