Mistral opens Large 4 weights as top-scoring open model built outside China

Open weights for a near-frontier model

Mistral has released downloadable weights for Large 4 (ML4), nicknamed "Le Chonk", a trillion-parameter model it had previewed on 6 October. The weights are available on Hugging Face under the Mistral Ecosystem Open-Weight License, which sets no revenue threshold. A separate commercial agreement is needed only when the user or an affiliate runs a model-as-a-service business, subject to specific exceptions. The model reads text and images, produces text and code, and activates 52 billion of its parameters for each task.

Benchmarks and the gap to open leaders

In its 29 October assessment, Artificial Analysis gives ML4 at high reasoning 26 points on Intelligence Index v5, up from 25 for the preview. That is one point above Muse Spark Open at xhigh, but six points behind the open-weight leaders DeepSeek V4.1 Flash 1017 and GLM-5.3, both tested at their maximum reasoning settings. On Terminal-Bench 4.0, which tests agentic coding in a terminal, ML4 rose from 28% at the preview to 41%, edging GLM-5.3 at 40%. On CyberGym-E2E, which measures whether a model can find, reproduce and patch memory-safety bugs in open-source C and C++ projects, ML4 scored 79%, level with MiMo-V2.6-Pro and ahead of DeepSeek V4.1 Flash at 76%.

ML4 at preview28%

ML4 at release41%

GLM-5.340%

ML4 at preview

28 %

ML4 at release

41 %

GLM-5.3

40 %

Data from the article text

ML479%

DeepSeek V4.1 Flash76%

ML4

79 %

DeepSeek V4.1 Flash

76 %

Data from the article text

Cost per task and coding efficiency

Artificial Analysis puts ML4's cost at $2.06 per Intelligence Index task, down from $5.54 for the preview, a 63% reduction. Its coding-agent assessment ranks ML4 second among current open-weight entries, while using 47% fewer tokens per task on average than GLM-5.3, a saving Mistral says compounds when agents run for hours. Early testers, described only as governments and enterprises, wanted more control over how long the model thinks, so Mistral added low and medium reasoning modes and made high mode 50% more efficient. The cost gains matter most for customers who want to keep control of their deployments.

ML4 preview5.54

ML4 release2.06

ML4 preview

5.54 USD per task

ML4 release

2.06 USD per task

Data from the article text

Marjorie Janiewicz, Mistral's chief revenue officer, told VentureBeat in an in-person interview that customers want models they can retain, customise around their own information and use without depending entirely on a provider's continuing permission. She put the core demand plainly:

“I want to own my model”

Orange Cyberdefense and data control

Orange Cyberdefense joined the Large 4 design partner programme on Thursday, 29 October. The French firm will introduce the model gradually, starting with technical audits and penetration testing, where specialists try to break into client systems to find weaknesses. Later uses include assessing how attackers could reach those systems and investigating breaches so staff can respond faster. Orange Cyberdefense will run its security models on computers in Mistral's European data centres, which it says helps clients keep control of their data. Financial terms were not disclosed, and the rollout initially targets business and public-sector clients.

Anis Trabelsi, head of innovation and AI at Orange Cyberdefense, said:

“This alliance between Orange Cyberdefense and Mistral AI strengthens the foundation of cybersecurity augmented by trustworthy AI”

Open weights as a strategic argument

Pierre Stock, Mistral's VP of science, told reporters at a press briefing:

“It's a pivotal moment in the global AI leadership race”

Stock argued that enterprises and governments need models that cannot be switched off remotely, comparing AI to electricity and warning against depending on a provider that could cut access at short notice. He said that once weights are published they spread across the internet, so access cannot be withdrawn. He also said there is no guarantee that Chinese models will avoid malicious behaviour, and that they can serve as threat vectors. Rivals argue that open models carry more risk. Mistral says it built ML4 with a fraction of the resources available to closed labs and Chinese rivals, a claim that is difficult to verify because the Chinese labs disclose little.