private-local.ai vs Private AI, now Limina: redaction or your own hardware

Published September 23, 2026Updated October 8, 202613 minutes read

private-local.ai vs Private AI, which has been called Limina since March 2026, is a choice between two different approaches to the same problem.

The problem is that certain material cannot simply be sent to a third party. Limina addresses it by removing the sensitive information from the material before the material is sent anywhere. We address it by helping you determine whether the model could instead run on a machine that you own. Removing the sensitive information is the smaller change, but the detection is automated, and artificial intelligence (AI) makes mistakes, so you cannot be fully certain that every sensitive item has been removed. A model on your own machine makes that question unnecessary.

When a machine of your own is the better choice

Redaction works when the sensitive information is a detail that the task can do without, such as a customer's name in a support request. It falls short in two situations. The first is material that must remain completely confidential, where a single undetected item would already be a breach, and by Limina's own benchmark some items are missed. The second is material whose value lies in the content itself, such as designs, source code, formulas or the terms of a transaction: this is not personal information, so redaction is not designed to remove it, and removing it would make the task impossible. In both situations, a model on a machine that your company owns is the better choice, because the material is processed inside your own network and does not reach any third party.

Industry	Scenario	Why redaction is not enough
Law firms	A lawyer has a model review a share purchase agreement and summarise the liability clauses for a client.	The parties, amounts and clauses are the substance of the review, so redaction would remove what the model needs. Professional secrecy also obliges the firm to keep client information confidential.
Tax advisers and auditors	A tax adviser has a model check a client's annual accounts and prepare notes for the tax return.	The figures, account numbers and company names are the content of the work, and a single overlooked detail would disclose a client's financial position to a third party.
Medical practices and clinics	A clinic has a model draft discharge letters from treatment notes.	Health data belongs to the special categories of personal data under the General Data Protection Regulation (GDPR). Even without names, a rare diagnosis combined with dates and places can identify a patient.
Manufacturing and engineering	An engineering department has a model compare design specifications, tolerances and process parameters across product generations.	The trade secret is the technical content itself, not personal information. Redaction is designed to find personal, medical and payment information, so it would pass the designs to the provider unchanged.
Software companies	A development team has a model review and restructure its proprietary source code.	The source code is the company's intellectual property and also the subject of the task, so it cannot be removed without removing the work itself.
Pharmaceutical and chemical research	A research team has a model analyse unpublished trial results and draft a patent application.	Research results and inventions are most valuable before publication or filing, and redaction does not recognise an invention as sensitive information.
Corporate finance and mergers	A finance team has a model analyse due diligence documents before a takeover is announced.	Before the announcement, the transaction can itself be inside information, and the company names, prices and dates cannot be removed without making the analysis meaningless.

private-local.ai vs Private AI (Limina): the main differences

Criterion	private-local.ai	Private AI, now Limina
What is provided	You receive temporary access to a named machine, model and serving configuration through a web address and an access key. The configuration is documented, so that a result can be reproduced later on equipment that you own.	Limina provides a software container that detects, redacts and replaces personal, health and payment information in text, documents, images, audio and structured records. It covers 52 languages and more than 50 categories of information.
Who operates the system	We operate the machine and the model for the duration of the test drive, and you connect your own tools and run your own workflows. If you then buy a machine, your company operates it.	Your organisation operates it. The container is designed to run on your own infrastructure, with one instance per machine, on Docker or Kubernetes. Limina also offers a hosted preview for evaluation before a purchase.
Where the processing takes place	During the test drive, the work runs on the machine that you selected, which we operate on your behalf. A machine that you buy afterwards runs wherever your company places it.	The processing takes place inside your own network or cloud account. Limina's documentation states that the data never leaves your infrastructure, and installation on a network without any external connection, known as an air-gapped installation, is supported.
Which models are used	The test drive uses open models that we name and fix to a specific version, selected to fit the hardware being tested. No proprietary model, whether ours or another company's, runs on the setup.	Limina provides no model, which is the point most often overlooked. It does not answer questions; it filters the material before it reaches the model that does, whether that model is hosted by a provider or runs on your own hardware.
How the costs are structured	The test drive is charged at an hourly rate, which is published on our pricing page. Afterwards, you decide whether to buy equipment.	Limina does not publish its prices. Licences for enterprises are negotiated individually.
Which organisations it suits	It suits companies with fewer than 5,000 employees that are considering the purchase of a machine and want to see it perform their own work first. Examples are a manufacturer that wants an assistant for its technical documentation, a law or tax firm that plans to process client files on its own hardware, and an automation agency whose clients ask for workflows that run locally.	It mainly suits large organisations in regulated industries, such as banks, insurers, pharmaceutical companies and hospital groups, that are satisfied with hosted models as long as no identifiable information reaches the provider. Its named customers include Boehringer Ingelheim, Zurich Insurance, MUFG Bank and Providence Health. An example is an insurer that wants a hosted model to summarise claims correspondence without passing on names, addresses or policy numbers.

What Limina is

Private AI is a company based in Toronto that was founded in 2019. On 5 March 2026, it renamed itself Limina; its original web address now forwards to the new one, while its documentation remains available at the old address. It sells a single product: a software container that finds sensitive information in your material and removes it before the material reaches a model, a data pipeline or an analyst.

Limina states that it recognises more than 50 categories of personal, medical and payment information in 52 languages, including text that changes language within a sentence, and that it processes plain text, documents, images, audio and structured records. Its principal claim is that it interprets context rather than matching fixed patterns, so that it can recognise a name even when the word does not look like one. According to its published figures, it misses between 0.2 and 7 per cent of sensitive items, compared with 13.8 to 46.5 per cent for the alternatives that it tested, on test sets of 45,000 words. By the company's own account, therefore, some sensitive items are not detected.

Limina's deployment model is simple, and it is informative for this comparison. Limina runs as one container per machine on Docker or Kubernetes, in your own cloud account or on your own premises, and it can also be installed on a network without any external connection. The published minimum for the version that runs on the processor alone is a standard 64-bit processor with 7.5 gigabytes of free memory and 50 gigabytes of disk space; the accelerated version requires an NVIDIA graphics card with at least 16 gigabytes of memory and is intended for organisations that make billions of requests per month. The modest hardware requirement is significant, because it means that Limina offers a route to data protection that does not begin with a capital purchase.

Limina's website names customers in regulated industries, among them Boehringer Ingelheim, Zurich Insurance, MUFG Bank and Providence Health. The company reports a System and Organization Controls 2 (SOC 2) audit and certification to the information-security standard ISO 27001, and states that it supports compliance with the Health Insurance Portability and Accountability Act (HIPAA), the General Data Protection Regulation (GDPR), the California Privacy Rights Act (CPRA) and the Payment Card Industry Data Security Standard (PCI DSS).

The web address private.ai is easily confused with Private AI, but it belongs to an unrelated consumer chat service.

Limina's strengths

If the only obstacle is that customer records must not reach a third party, removing the identifying information is a smaller change than moving the model to your own hardware.
It runs on ordinary hardware and on networks without any external connection, so meeting a data-protection requirement needs neither a capital purchase nor a procurement process.
It names large customers in regulated industries and reports a SOC 2 audit and ISO 27001 certification.

Limina's limitations

Limina is neither a model nor an alternative to one. A model must still answer your questions, and that model may still run at a cloud provider, with the corresponding costs and dependence.
Redaction protects personal information, not trade secrets: designs, source code and formulas reach the model provider unchanged. Work that relies on the redacted details, such as a contract review that depends on who signed which document, is poorly served once the names have been removed.
The detection is automated, and by Limina's own benchmark it misses some sensitive items, so you cannot be fully certain that everything has been removed. Limina also does not publish its prices.

Which option suits which organisation

Limina is the better choice if

The only obstacle to using ChatGPT or a similar AI chatbot is that identifiable information must not reach it, and an occasional undetected item would be acceptable.
You are satisfied with the quality of hosted models, and their cost is not a concern.
You need a data-protection solution quickly, on hardware that you already own, with certifications that an auditor will recognise.

A machine of your own is the better choice if

Your material requires complete data privacy. By Limina's own benchmark, redaction misses between 0.2 and 7 per cent of sensitive items, which is far too high for client files, patient records or trade secrets, where a single undetected item is already a breach. Redaction also fails when the work depends on the details that would be removed, or when the material is your intellectual property, such as designs, source code or formulas.
Your main concern is cost or volume rather than data protection: the same task runs continuously, and a subscription per person is no longer economical.
You want to reduce your dependence on a model provider rather than only protect the material that you send to one.

Sources and notes

All information about Limina on this page was read from the company's own website and documentation on 22 September 2026 and checked again on 23 September 2026, including the date of the renaming, the numbers of languages and categories, the published minimum system requirements and the accuracy figures. These are Limina's own claims and benchmark results, and they are reported as such.

The accuracy figures should be read with caution. Limina chose the test sets, the competitors for comparison and the definition of a missed item, and a benchmark published by a vendor is produced by a party with an interest in the result. This does not make the figures wrong, but we are not aware of any independent verification of them.

If any information on this page is out of date, please let us know and we will correct it.

Further reading: the other four comparisons cover ChatGPT, Palantir, the open-source engine LocalAI and the on-premises assistant Seesa, and the pricing page states the hourly rate.

Questions about private-local.ai and Private AI

private-local.ai vs Private AI: which one should we choose?

That depends on what currently prevents you from using a hosted model. If the only obstacle is that identifiable information must not reach a third party, and an occasional undetected item is acceptable, redaction with Private AI, now Limina, is the smaller change. If the obstacle is the cost of large volumes, dependence on a provider or material that cannot be made safe by removing names from it, the question is whether a model on your own hardware could do the work, and a test drive with a representative task answers it.

Is Private AI the same company as Limina?

Yes. Private AI, founded in Toronto in 2019, renamed itself Limina on 5 March 2026; its original web address now forwards to the new one, while its documentation remains available at the old address. Both names refer to the same company and the same product. Two other names are easily confused with it: the web address private.ai belongs to an unrelated consumer chat subscription run by a different company, and the name PrivateGPT has been used both by Private AI and by an unrelated open-source project.

Does Limina run a language model for us?

No, and this is the most common misunderstanding. Limina is a filter, not a model: it finds personal, medical and payment information in your material and removes or replaces it before the material is passed on. A separate model must still answer your questions, and that model may well be a hosted service, with the corresponding costs and dependence.

Is redaction enough if our data must remain completely private?

No. By Limina's own benchmark, redaction misses between 0.2 and 7 per cent of sensitive items, so in documents containing 1,000 names, amounts or diagnoses, between 2 and 70 of them would reach the model provider undetected. For client files, patient records or trade secrets, where a single undetected item is already a breach, that rate is far too high. A model on a machine that your company owns avoids the problem, because the material is processed inside your own network and does not reach any third party.

Find out whether a machine of your own suits your work

Choose a setup, connect the tool that you already use and run a representative task. Within one or two hours, you will know whether a machine of your own can do the work, before you buy one.

Request an endpoint