Mistral Large 4 weights released with high cybersecurity reproduction scores
The open-weight multimodal model achieves 82% on vulnerability reproduction tasks where closed systems often refuse to engage.
Key points
- Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test for reproducing and patching vulnerabilities.
- The model features 1 trillion total parameters with 49 billion active, supporting a context window of one million tokens.
- It is distributed under the Mistral Ecosystem Open-Weight License, which restricts certain model-as-a-service offerings.
Mistral Large 4, also known as Le Chonk, is now available as an open-weight multimodal model on Hugging Face, according to an official blog post from the platform.reported The release highlights the model’s ability to reproduce real software vulnerabilities and subsequently patch them, a capability that several closed frontier models reportedly refuse to perform. This distinction is significant for security defenders who require functional reproduction of flaws before remediation can occur.
How we got here
Previous iterations of large language models often relied on closed APIs for advanced reasoning tasks, which introduced friction for security teams. According to the Hugging Face post, closed-source filters frequently block inputs that resemble attacks, even when those inputs are part of legitimate defensive research.reported This behavior can disrupt incident response workflows. Furthermore, API-dependent teams risk losing access during critical moments. Open weights offer a solution by placing deployment and usage policy control directly with the user.
What is new
The newly released Mistral Large 4 combines text and image inputs with a Mixture-of-Experts architecture.confirmed It uses sparse attention mechanisms to select relevant experts and context positions dynamically. The model achieves an 82% score on the Artificial Analysis Cyber Index test, which requires reproducing open-source vulnerabilities and applying patches. Mistral reports that closed models score near zero on this specific metric because they decline to attempt the reproduction step entirely.reported
What to watch
Developers may use the open weights to fine-tune smaller models for specific security tasks using traces from Mistral Large 4.reported The post notes that reproduce-then-patch tasks provide verifiable rewards, making them suitable for reinforcement learning environments. Researchers can also directly measure the model’s refusal behaviour on benchmark suites like JailbreakBench and AgentHarm. The model maintains a higher average refusal rate for harmful requests than other open-weight alternatives, suggesting robust safety filters remain in place.
Availability and limits
Mistral Large 4 is accessible via Hugging Face, with additional GGUF formats provided by Unsloth and the llama.cpp team.reported The model operates under the Mistral Ecosystem Open-Weight License, an MIT derivative that imposes specific limitations on model-as-a-service offerings. Users must adhere to these terms when deploying the model commercially or as an API. The architecture includes a Pixtral vision encoder and supports multilingual text generation. No additional cost is mentioned for accessing the base weights, but compliance with the license is mandatory for all downstream applications.
Why it matters
- Security teams can locally deploy a model that reproduces vulnerabilities, bypassing the refusal rates common in closed API systems during defensive analysis.
- The open-weight nature allows developers to audit refusal behaviour directly, rather than relying on vendor descriptions, enhancing transparency in security tooling.
- The high parameter count and context window enable complex multimodal security tasks, such as analysing crash logs and malware samples within a single session.
- The licensing terms restrict certain commercial service models, requiring developers to carefully review compliance before building model-as-a-service applications.
Topics: Mistral, cybersecurity, open weights, vulnerability research, multimodal models
Updated
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Ask the article
Local keyword search of this article; no generated answers.Which of these is confirmed?
Check the answer
The newly released Mistral Large 4 combines text and image inputs with a Mixture-of-Experts architecture.
The matched primary source and announcement wording support the confirmed label.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
What is the performance of Mistral Large 4 on cybersecurity benchmarks?
It scores 82% on the Artificial Analysis Cyber Index test for reproducing and patching vulnerabilities.
Can Mistral Large 4 process images as well as text?
Yes, it is a multimodal model that takes both text and images as input to generate text outputs.
What license governs the use of Mistral Large 4 weights?
It is distributed under the Mistral Ecosystem Open-Weight License, which limits certain model-as-a-service offerings.
Connected stories
Reading receipt
- Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test for reproducing and patching vulnerabilities.
- The model features 1 trillion total parameters with 49 billion active, supporting a context window of one million tokens.
- It is distributed under the Mistral Ecosystem Open-Weight License, which restricts certain model-as-a-service offerings.