September 22, 2026 — Toronto, Canada
Today, we are releasing Vinci Cyber, a family of three open-weight models built for defensive cybersecurity: 8B, 30B, and 123B parameters.
All three models were fine-tuned in Canada. The 8B and 30B models are released under Apache 2.0; the 123B model uses Mistral’s Modified MIT licence.
They are available to developers, researchers, businesses, and security teams in Canada and around the world, subject to each model’s licence terms. [5–8]
This is our first cybersecurity model family at Vinci.
And it exists because we think defenders need more control over the intelligence they depend on.
Cybersecurity is becoming an AI problem
The cybersecurity environment is getting harder, not easier.
Canada’s Centre for Cyber Security assesses that ransomware incidents in Canada are rising overall, with annual increases across most sectors.
It identifies ransomware as the leading cybercrime threat to Canada’s critical infrastructure. [1, 2]
The trend is global. Ransomware appeared in 48% of breaches in Verizon’s 2026 Data Breach Investigations Report dataset.
Vulnerability exploitation overtook stolen credentials as the most common initial entry point, accounting for 31% of breaches.
Its incident window runs from November 1, 2024, through October 31, 2025. [3]
IBM’s 2026 X-Force Threat Intelligence Index reported a 49% year-over-year increase in active ransomware and extortion groups and a 44% increase in attacks beginning with exploitation of public-facing applications. The first number counts groups, not attacks or victims. [4]
At the same time, attackers are getting access to better automation and increasingly capable AI.
The Cyber Centre and IBM both describe the role of AI in lowering barriers and accelerating malicious activity. [1, 4]
Defenders need better tools too.
Why open-weight cybersecurity models?
Some of the world’s strongest AI systems are produced by large frontier labs in the United States and elsewhere.
We use many of them ourselves, and the capabilities they have developed are remarkable.
But cybersecurity is a particularly difficult domain for closed AI systems.
Providers have legitimate reasons to place safeguards around cyber capabilities. A model capable of finding vulnerabilities can potentially help someone exploit them. Access reviews, usage policies, rate limits, and additional controls are therefore understandable.
We experienced those constraints ourselves while building Vinci.
But they revealed another problem.
A security team protecting its own infrastructure may need to analyze millions of lines of proprietary code. A bank may not want security telemetry leaving its environment.
A government may require models to run entirely inside infrastructure it controls. A security researcher may need reproducible access to a model rather than an API whose behaviour, limits, or availability can change.
For those organizations, control is itself a security property.
They should be able to run models on infrastructure they control, inspect how they were evaluated, integrate them into their own security systems, and continue operating without depending on permission from a remote inference provider.
That is the idea behind Vinci Cyber. Self-hosting creates control over deployment; it does not by itself guarantee secure operation or correct model output.
Deployment under an organization’s control: a conceptual illustration, not a system diagram or evaluation result.
Three models, for different environments
We are releasing three sizes of Vinci Cyber.
Vinci Cyber 8B 1.0 is the smallest model in the family, intended for teams weighing deployment cost, latency, and hardware requirements.
Vinci Cyber 30B 1.0 offers a larger model for organizations operating their own GPU infrastructure. Actual latency, memory requirements, and throughput depend on the serving configuration and workload.
Vinci Cyber 123B 1.0 is our largest model, fine-tuned from Mistral’s Devstral 2 123B foundation. Its size is not evidence of better security outcomes; it needs to be assessed on its own evaluation record. [8, 9]
The 8B and 30B models are fine-tuned from IBM Granite 4.1 foundations; 123B is fine-tuned from Mistral’s Devstral 2. These are Canadian-developed specializations of existing foundations, not models pretrained from scratch by Vinci. [5, 6, 9]
Vinci Cyber 8B 1.0, 30B 1.0, and 123B 1.0. The artwork identifies the family; it is not a performance comparison.
One practical licence distinction: 8B and 30B use Apache 2.0. For 123B, Mistral’s Modified MIT LICENSE does not authorize use under that licence if your company’s - or your employer’s - global consolidated revenue exceeded US$20 million in the preceding month.
That restriction also covers derivatives. US$240 million is only an annualized comparison, not an annual-revenue eligibility test.
Above the monthly threshold, contact Mistral about a separate commercial licence. Read the accompanying terms before deployment. [7, 8]
Our broader direction is defensive cybersecurity: secure-code analysis, vulnerability reasoning, remediation, and security analysis.
The current 8B and 30B model cards make a narrower, concrete claim: bounded Terraform and OpenTofu infrastructure-as-code work, including repair proposals, no-change decisions, and recovery after an unsuccessful edit.
That evidence should not be read as proof of general application-code repair or broad cybersecurity superiority. [5, 6]
Rather than putting every technical result into this announcement, we are publishing the detailed evaluation record alongside the models.
That includes benchmark results, methodology, limitations, and provenance so that researchers and security teams can examine the evidence themselves.
Each model must be assessed on its own record; results do not automatically transfer between sizes or to quantized formats.
We think that distinction matters.
The blog post tells you why we built the models. The model cards should help you decide whether—and for which tasks—you should trust them.
Built in Toronto
We started Vinci at the end of June 2026 with a simple idea: Canada should be capable of building important AI systems of its own.
Not because every model has to be Canadian.
And not because technological independence means technological isolation.
Modern AI is built on an enormous international ecosystem of research, open models, software, chips, data centres, researchers, and companies. We benefit enormously from that ecosystem.
But Canada should also contribute to it.
We want Canadian researchers and companies to train models, publish research, operate compute infrastructure, build tools, and develop capabilities that are useful well beyond our borders.
Vinci Cyber is one small step in that direction.
We are a small team in Toronto. We fine-tune our models on Canadian infrastructure. We are building in the open wherever we reasonably can. And we intend to keep releasing what we learn.
Open weights do not mean every part of the training process is public. The 8B and 30B specialization corpus is private; Apache 2.0 applies to the released weights, not that private corpus. [5, 6]
Defensive AI should not be scarce
There is a real tension at the frontier of cybersecurity AI.
Making increasingly capable models available creates genuine security questions. Pretending otherwise would be irresponsible.
But making advanced defensive capabilities scarce creates risks too.
Attackers do not need enterprise procurement processes. They do not need security reviews from an AI vendor.
Cybercrime markets provide access to automated tooling, stolen infrastructure, and ransomware-as-a-service, while AI is lowering some barriers to entry. [1, 4]
The answer cannot be to leave legitimate defenders with weaker tools.
Our approach is to build models specifically around defensive cybersecurity, evaluate them, document what they can and cannot do, and make the weights available so organizations can deploy them under their own control.
Deployment still requires independent validation, access controls, and human review. A proposed patch is not a verified repair, and clearing a scanner finding is not proof that a security property has been restored. [5, 6]
Vinci Cyber 8B 1.0, Vinci Cyber 30B 1.0, and Vinci Cyber 123B 1.0 are the beginning of that work.
The weights, model cards, evaluations, and technical documentation are available beginning today through SimpleDirect on Hugging Face: Vinci Cyber 8B 1.0, Vinci Cyber 30B 1.0, and Vinci Cyber 123B 1.0.
We hope researchers test them.
We hope security teams break them.
And we hope people build better defenses with them.
— George Pu
Founder, Vinci
Sources and technical references
- Canadian Centre for Cyber Security, Ransomware Threat Outlook 2025–2027. Supports the Canadian trend, ransomware ecosystem, and AI threat context. Assessment based on information available through September 4, 2025; incidents are underreported.
- Canadian Centre for Cyber Security, National Cyber Threat Assessment 2025–2026. Supports the assessment of ransomware as the leading cybercrime threat to Canadian critical infrastructure.
- Verizon, 2026 Data Breach Investigations Report, released May 19, 2026, official findings announcement, and report overview and methodology FAQ. Ransomware prevalence is reported on printed page 11. The percentages describe breaches in the report’s dataset, not all organizations; its incident window is November 1, 2024–October 31, 2025.
- IBM, 2026 X-Force Threat Intelligence Index: official findings, February 25, 2026. The 49% figure concerns active ransomware and extortion groups; the 44% figure concerns growth in attacks beginning with exploitation of public-facing applications.
- SimpleDirect, Vinci Cyber 8B 1.0 model card. Primary reference for lineage, scope, evaluation limitations, training-data disclosure, and licence.
- SimpleDirect, Vinci Cyber 30B 1.0 model card. Primary reference for this model’s own evaluation record; it must not be substituted for the 8B or 123B record.
- Apache License 2.0. Applies to the 8B and 30B released weights, not the 123B model. Consult each artifact’s accompanying LICENSE and NOTICE files for its release terms.
- Mistral AI, Devstral 2 123B Modified MIT LICENSE. Primary source for the preceding-month global consolidated revenue restriction, its application to derivatives, and the separate commercial-licence option.
- Mistral AI, Devstral-2-123B-Instruct-2512 model card, and SimpleDirect, Vinci Cyber 123B 1.0 model card. References for the parent foundation and the 123B release’s own provenance and limitations; the 8B and 30B results are not evidence for 123B.
The external threat reports provide context; they are not evaluations or endorsements of Vinci Cyber.
Statements about Vinci’s origin, Canadian fine-tuning, experience, and release plans are first-party statements by the author.
