The Information Machine
ConcludedRan 6 daysin the edition since 31 Oct 20263 sourcesupdated 5 Nov 2026

Mistral Large 4 open weights release

The gist

Latest: Mistral Large 4 open weights released with 1T parameters, API prices cut

A frontier-scale model available for self-hosting lets organizations with strict data-handling requirements run inference without sending data to an external API. The safety-driven split release, withholding a cyber-capable variant from public access, sets a visible precedent for how labs can handle dual-use capability concerns at open-weight release.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

Mistral released downloadable weights for Large 4 on Hugging Face, making the model available for self-hosting. The model is a mixture-of-experts architecture with 1.05 trillion total parameters and roughly 49 billion active at any inference step, with a 1 million token context window. API pricing on Mistral's service dropped to $0.70, $0.07, and $2.10 per million input, cached, and output tokens, from prior prices of $1.36, $0.14, and $4.18. The release uses the Mistral Ecosystem Open-Weight license rather than the Apache 2.0 license that covered Large 3; internal self-hosting is allowed but selling the model as a service requires a separate commercial license.

Mistral split the release on safety grounds. Internal testing suggested Large 4 could increase attackers' real-world abilities, even though the model refused 96.6 percent of malicious cyber requests on selected benchmarks. A stronger cyber-specialized variant was withheld from public release and provided only to vetted partners such as Orange Cyberdefense for authorized penetration testing.

Mistral's own benchmark claims show three weeks of reinforcement learning after an earlier preview lifted the Terminal-Bench 4.0 score from 28 to 41 percent, one point above GLM-5.3, and put Large 4 in a tie with GLM-5.3 on Artificial Analysis's Coding Agent Index at roughly 48 percent while consuming 47 percent fewer tokens per task. Third-party evaluators ThibaultJaigu and RequestyAI placed it as the second-best open-source model on their coding benchmark.

Mistral has published several capability demonstrations. In one, given a short prompt, Large 4 wrote a quantum chemistry self-consistent field solver from scratch, validated its output, and built a 3D visualization; Mistral said the model independently chose how to break the molecule and which orbitals to display. In a second, the model built a real-time 3D tracking application for roughly 19,000 satellites and debris fragments in Earth orbit, called Orbital Trace; Mistral reported the model built the app, tested it from a user perspective, optimized for performance, and passed strict verification. In a third demonstration, Mistral tested Large 4 against a recent Django CVE, deploying 133 agents to map the codebase, identify flaws, and test attack scenarios; the model confirmed the bug and produced a patch, with the full triage-to-patch workflow shown at 90x speed.

How it developed
4 November 2026in the 5 Nov edition

Mistral published a demonstration of Large 4 against a real Django CVE, deploying 133 agents to triage the vulnerability, reproduce it, and generate a patch, with the full workflow shown at 90x speed.

The test extends a series that has included a quantum chemistry solver and a satellite tracker, and sits in a domain adjacent to the stronger cyber variant Mistral withheld from public release, restricting it to vetted partners such as Orange Cyberdefense for authorized penetration testing.

First citedMistral on X
3 November 2026in the 4 Nov edition

Mistral published a second capability demonstration for Large 4, a 1.05-trillion-parameter mixture-of-experts model: the model built Orbital Trace, a real-time 3D tracker of approximately 19,000 satellites and debris fragments, handling testing, optimization, and verification autonomously according to Mistral.

Mistral also published a technical report on AlphaXiv covering the model's development backstory and described the downloadable weights as more efficient and compact than the earlier Preview version.

First citedMistral on X
2 November 2026in the 3 Nov edition

On November 2, Mistral demonstrated Large 4, a 1-trillion-parameter mixture-of-experts model released October 29, writing a quantum chemistry self-consistent field solver from scratch, validating its own output, and producing a 3D molecular visualization.

According to Mistral, the model independently chose how to break the molecule and which orbitals to display without further input.

1 November 2026

Downloadable weights confirmed available for self-hosting on own infrastructure

First in the edition · 31 October 2026
30 October 2026in the 31 Oct edition

On October 29, Mistral published Large 4, a 1-trillion-parameter mixture-of-experts model, on Hugging Face with API prices cut roughly 50 percent.

Three weeks of reinforcement learning added after a preview raised its Terminal-Bench 4.0 score from 28 to 41 percent, and ThibaultJaigu and RequestyAI ranked it second among open-source models on their coding benchmark. The release license requires Model-as-a-Service providers to obtain a separate commercial license, unlike Large 3's Apache 2.0 terms, and Mistral withheld a stronger cyber variant, giving it only to vetted partners such as Orange Cyberdefense after finding it could raise attackers' real-world capabilities.

28 October 2026

Mistral launched the Responses API in public preview, designed for building AI agents, compatible with all Mistral-hosted chat models.

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy