Skip to content
Sign in
Technology · Global

Mistral introduces Large 4 with 1.05 trillion parameters

Mistral unveiled Large 4, a model with roughly 1.05 trillion parameters and a mixture-of-experts design. HowToGeek says 52 billion parameters are active at a time and reports the company’s claims about the model’s performance in software security testing and other fields.

AI-assisted synthesis · Based on 4 outlets' reporting · Article text read from 1 of 4 outletsHow it's made

View timeline · 7 entries
4 outletsLatest reporting 3 min read
Coverage: Too few ratedL 0 · C 2 · R 0 · 2 unrated

What happened

Key facts

  • Large 4 has roughly 1.05 trillion parameters; shattered.io’s headline describes it as a 1T-parameter model (per howtogeek.com).
  • Mistral said a preview version of Large 4 scored 82% on an evaluation requiring models to reproduce and patch real software vulnerabilities (per howtogeek.com).
  • Mistral said Large 4 performs strongly in finance and legal work, scientific coding and visual understanding (per howtogeek.com).
  • inkl’s headline says Mistral calls Large 4 the best open model for agentic coding and cyber defense (per inkl.com, howtogeek.com, shattered.io).

The brief

Mistral has introduced Large 4, an open-weight model with roughly 1.05 trillion parameters, positioning it for coding, complex reasoning, image analysis and business automation. HowToGeek reports that the model uses a mixture-of-experts architecture, with 52 billion parameters active at a time.

The outlet describes Large 4 as the company’s latest effort to compete with major AI developers in the United States and China. It says Mistral previously presented the model as suited to a range of professional and technical tasks, including finance, legal work, scientific coding and visual understanding.

Those assessments of its performance come from Mistral, as reported by HowToGeek. Mistral said during a preview that Large 4 scored 82% on an evaluation in which models had to reproduce and patch real software vulnerabilities. HowToGeek says this was the highest score reported on that evaluation.

The account does not identify the evaluation by name or provide competing scores, so the result is presented as the company’s stated performance claim rather than an independently established comparison.

The headlines from other outlets provide limited additional claims. shattered.io’s headline describes Large 4 as a 1T-parameter model and cites a 61.7% DeepSWE figure, without explaining the metric in the headline. inkl’s headline says Mistral calls the model the best open model for agentic coding and cyber defense.

These headline descriptions do not provide further detail about the tests or criteria behind those characterizations. HowToGeek also places the launch in the context of Mistral’s previous products and financing. It says the company moved from smaller open models to larger systems such as Mistral Large 3, and reports that Mistral raised €3 billion in September 2026.

The outlet characterizes Large 4 as a renewed effort to compete with leading U.S. and Chinese AI developers.

The model’s stated scope spans software development, reasoning, image analysis and business automation, while the performance figures described in the coverage come from different claims: Mistral’s reported 82% vulnerability-testing result and shattered.io’s headline figure for DeepSWE.

The available descriptions do not explain how those two measures relate or whether they assess the same capability.

Why it matters

  • Large 4’s mixture-of-experts design combines roughly 1.05 trillion total parameters with 52 billion active at a time, the model’s reported technical configuration.
  • Mistral says the model is intended for coding, complex reasoning, image analysis and business automation, spanning technical and business uses.
  • Mistral’s reported 82% vulnerability-testing score is a company performance claim; shattered.io’s headline separately cites 61.7% for DeepSWE, a figure whose metric is not explained there.
  • HowToGeek reports that Mistral raised €3 billion in September 2026 and frames Large 4 as part of its effort to compete with AI developers in the United States and China.

What to watch next

  • Whether Mistral provides further performance results for Large 4 on the software-vulnerability evaluation it described.
  • How Mistral substantiates its claims about Large 4 in finance, legal work, scientific coding and visual understanding.
  • What metric shattered.io’s cited 61.7% DeepSWE figure measures and how it relates to the model’s other reported results.

How it's being covered

Coverage analysisShow details
5 source articles · 4 unique outlets
Outlet political mix: Too few rated· 2 Center · 2 unrated
Where the outlets are based: Europe 50% · Unknown 50%
Outlet political mix
Too few rated
Left
Center
Right
0 outlets
2 outlets
0 outlets
The marker shows coverage balance: each unique outlet counts once. 2 unrated outlets are not included.
Geography
Where the outlets covering this story are based.
4 unique outletsDominant coverage: Europe
How we measure coverage

How outlets' accounts compare

Updat3 does not yet have enough full reports to compare how outlets framed this story.

Timeline

Developments from this story's sources, dated background we compiled and fact-checked, and our earlier coverage. Each row says which.
  1. The Paris-based company was founded in 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix, former researchers at Google DeepMind and Meta.
    Background · from our historical context
  2. Mistral released Mistral 7B under the Apache 2.0 license.
    Background · compiled and fact-checked by Updat3
  3. Mistral released Mixtral 8x7B, a mixture-of-experts model.
    Background · compiled and fact-checked by Updat3
  4. On February 26, 2024, Mistral released its first Mistral Large model and presented it as a competitor to leading U.S. systems.
    Background · from our historical context
  5. Mistral released Mistral Large 2.
    Background · compiled and fact-checked by Updat3
  6. Mistral’s immediate backdrop was its move from smaller open models to Mistral Large 3 in late 2025.
    Background · from our historical context
  7. In September 2026, the company raised €3 billion, according to How-To Geek, adding financial backing before its next major model announcement.
    Background · from our historical context

How we got here

Mistral’s immediate backdrop was its move from smaller open models to Mistral Large 3 in late 2025. In September 2026, the company raised €3 billion, according to How-To Geek, adding financial backing before its next major model announcement.

The Paris-based company was founded in 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix, former researchers at Google DeepMind and Meta. It introduced Mistral 7B, an openly released 7-billion-parameter model, in September 2023, then published Mixtral 8x7B in December 2023, using a mixture-of-experts design that activates only part of a model for each task.

Original sources

5 articles from 4 outlets

Every story links to the reporting it was built from. How we choose sources

Spotted a factual error, a broken link or a misleading summary? Tell us.