AICanRun
AICanRun is a free web tool that checks which local AI models fit your exact phone or Mac, showing transparent memory math, estimated speed, the recommended quantization and a compatible beginner app. No account, nothing to install.
What is AICanRun?
AICanRun is a free compatibility service built around one narrow question: which local AI models will actually run on the phone or Mac you already own? Instead of generic hardware advice, it holds a database of 138 phones across sixteen brands and 38 exact Mac configurations, each described by chip, memory tier and memory bandwidth, and matches them against 75 models available in 462 quantizations.
The site works in three stages it labels Detect, Match and Run. You identify your device, the service performs its memory arithmetic — model weights, KV cache and runtime overhead, measured against what remains once the operating system takes its share — and it returns a verdict together with the quantization to download and the application able to open that file.
Its sitemap holds roughly 2,686 pages, most of them a single model-and-device pair. A verdict page states whether the model fits, the recommended quant, the gigabytes required against those available, an estimated decoding speed in tokens per second, and how long a 300-word reply should take. For Bonsai 27B at one bit on an iPhone 17 Pro Max, it reports a 3.8 GB download using about 4.8 GB of the 7.8 GB a model can address, at roughly 9.1 tokens per second.
Two decisions set it apart. First, hardware fit and application support are checked separately: a file can fit in memory and still lack a consumer app able to open it, so the site names PocketPal, LM Studio, Locally AI, Jan, Ollama or Msty according to the format, and refuses to pass a terminal command off as a beginner path. Second, it labels its own confidence — Estimated figures come from a published formula, Measured ones from reviewed real-hardware evidence, and verified ones from its own testing protocol.
Around the checker sit a glossary, a blog of tutorials and comparisons, per-model requirement pages and a published benchmark methodology. The site presents itself as an independent project built by a small group of local AI enthusiasts, unaffiliated with device makers, model publishers or app developers. It is free, needs no account and installs nothing.
What it does
- Tell you whether a given local AI model fits your exact phone or Mac configuration
- Show the memory arithmetic behind the verdict, weights, KV cache and runtime overhead included
- Estimate decoding speed in tokens per second, and how long a 300-word answer would take
- Recommend the quantization to download for your device
- Name a consumer application that can actually open that file today
- Walk you from installing the app to a first offline prompt
- Compare 75 models and 462 quantizations across 138 phones and 38 Mac configurations
When to use AICanRun / When not to
A quick filter to help you decide if AICanRun is the right fit.
When to use AICanRun
- Beginners who want to run an AI model offline but have no idea where to start
- Owners of a recent phone or Mac who do not know what their memory tier actually allows
- Privacy-minded users who want AI without a subscription and without an internet connection
- Developers and product managers scoping whether on-device AI is realistic on a target fleet
- Buyers comparing unified-memory tiers before choosing a Mac configuration
When not to use AICanRun
- Anyone expecting the site to run the model: it advises, it does not host or execute anything
- Engineers who need measured benchmarks, since most published figures are formula estimates
- Users looking for a cloud assistant, an API or an integration with their existing stack
- Non-English speakers, as the interface and every page are English-only
- Anyone wanting a guarantee before a hardware purchase, which the site explicitly declines to give
How to use AICanRun
A typical end-to-end flow, from setup to results.
- Open the site and go to the checker; no account, payment or installation is required
- Choose the Phone tab or the Mac tab depending on the device you want to test
- For a phone, let the site detect it in one tap or pick it from one of sixteen brands
- For a Mac, select the exact chip and the exact unified-memory tier, which changes the verdict
- Read the returned list of models, with grade, parameters, quant size, memory needed and estimated speed
- Open a model to see its dedicated verdict page for your device
- Note the recommended quantization and the gigabytes it will occupy
- Check which consumer application the page names for that exact file format
- Follow the step-by-step path from installing that app to sending a first offline message
- Verify the model still answers with the network switched off
Pros & Cons
Pros
- Entirely free, with no account, no payment and nothing to install
- Shows its calculation instead of handing down an opaque verdict
- Labels estimates as estimates and never presents them as measurements
- Separates hardware fit from application support, a distinction most sources skip
- Granular coverage down to the exact memory tier, not just the device model
- Beginner-oriented by design, with no terminal command disguised as a consumer path
- Publishes a reproducible benchmark protocol that anyone can challenge
Cons
- Most published figures are formula estimates rather than measurements on real hardware
- No privacy policy, no terms and conditions and no legal notice anywhere on the site
- Anonymous operator: no company name, no postal address, no named founder
- Very young project, with the domain registered in July 2026 and no archived history
- English only, with no language selector
- No API, no integration and no mobile application
- Nothing is saved between visits, since the site has no account system
Pricing & Plans
The service is entirely free. There is no paid tier, no usage quota and no paywall, and the site publishes no pricing page. No account or payment method is required at any point, so there is no lowest price point to report. The operator states that any affiliate links or sponsored placements added in future would be clearly labelled and would never alter a hardware or app-support verdict.
Data, GDPR & hosting
A consolidated view of how AICanRun handles your data.
GDPR overview
There is no GDPR mention of any kind on this site. AICanRun publishes no privacy policy, no terms and conditions and no legal notice: the paths /privacy, /privacy-policy, /terms and /legal all return 404. No data processing agreement is offered, no subprocessors are listed, no Article 27 EU representative is designated and no data protection officer is named. No cookie or analytics disclosure appears anywhere in the collected pages. The only published contact is a general address, hey@aicanrun.com, rather than a privacy channel, and the operator is identified merely as a small group of local AI enthusiasts, with no company name, address or country. This silence prevents any conclusion in either direction: the site neither claims compliance nor denies it. Readers with a regulatory requirement should treat the matter as open and ask the operator directly.
Who owns the data?
AICanRun publishes no terms and conditions and no privacy policy: the paths /terms, /terms-of-service, /legal and /privacy all return 404. There is therefore no contractual statement about who owns any data, what may be done with it, or by whom. In practice the site offers no account, no login and no form, so an ordinary visit transfers no personal data the operator could claim. The only voluntary transfer is an email sent to hey@aicanrun.com when a reader submits a benchmark result; the site says such submissions are reviewed before publication, but sets no ownership terms around them. Anyone needing a contractual guarantee should ask before sending anything.
Reuse rights
No terms of use exist, so the site grants no explicit licence and states no explicit restriction on reusing what it publishes. Its compatibility verdicts, memory calculations, speed estimates and device specifications are freely readable without an account, but nothing says whether they may be republished, quoted at length or used commercially. Ordinary copyright applies by default, and the footer carries a 2026 copyright notice in the AICanRun name. Readers who submit a benchmark by email receive no statement about how their contribution will be credited or licensed, beyond the note that submissions are reviewed before they can appear as measured results. Written permission should be sought for any reuse beyond normal citation.
Data retention & training
Hosting summary
AICanRun publishes nothing about data hosting. There is no privacy policy, no terms and no trust or security page, so no jurisdiction, region, provider or safeguard is disclosed. Technical observation of the infrastructure, which is not a commitment by the operator, shows the domain resolving to 64.29.17.1, an anycast address announced within AS16509 (Amazon), served behind an edge platform and registered through Cloudflare. Those are the ordinary components of a static site delivered from a global content network and say nothing about where user data would be stored. In practice the question is largely moot: the service requires no account, no login and no form, so a visit creates no user record for the operator to host. The single exception is a benchmark result sent voluntarily by email, which lands in an ordinary mailbox under unstated conditions. Readers with hosting or residency requirements will find nothing here to rely on and should ask before sending anything.
Things to keep in mind
Risks and trade-offs to weigh before adopting AICanRun.
- Nearly every figure is an estimate, not a measurement, and that nuance is easy to forget once a number looks precise
- The site itself warns against basing a hardware purchase on estimates alone, yet the temptation to do so is real
- Specifications for unreleased devices are flagged as rumoured and may simply be wrong
- Model files, runtimes and app support change quickly, so a verdict can go stale between visits
- No privacy policy, terms or legal notice exist, leaving no recourse and no stated commitment of any kind
- The operator is anonymous and the project only weeks old, so there is no track record to judge it by
- A free service with no visible business model carries no guarantee of remaining online or maintained
Setup & Integrations
Technical difficulty
Using the site itself is effortless: no account, no installation, no payment, and a verdict in two clicks. The difficulty sits downstream, in what it recommends. Running a local model still means downloading a file of several gigabytes and installing a third-party application, and grasping a few terms — quantization, usable memory, context length, tokens per second — which the glossary exists to explain. AICanRun deliberately targets beginners, favours point-and-click applications and never disguises a terminal command as a consumer path, so the route stays realistic for a patient non-specialist.
Deployment
Supported languages
Behind AICanRun
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Does AICanRun cost anything or require an account?
Does the site actually run the AI model for me?
Are the speed figures real benchmarks?
Can I run AI without an internet connection?
How much memory do I actually need?
Which application will it recommend?
Which devices and models are covered?
Can I submit a measurement of my own?
Who publishes the site?
Can I rely on it before buying a phone or a Mac?
Should you pick AICanRun?
AICanRun does one thing, and does it with unusual candour. The question it answers — will this model actually run on the device in my hand — is genuinely hard for a newcomer, and the usual answers are either vague or wrong. By keying on exact memory tiers rather than device names, showing its arithmetic, and separating whether a file fits from whether any application can open it, the site gives a beginner something to act on instead of a shrug.
Its most valuable habit is refusing to overclaim. Nearly every figure is a formula estimate, and it says so on the page rather than in a footnote, reserving measured and verified for evidence it has actually reviewed. It publishes its benchmark protocol so the numbers can be challenged, and warns in its own footer against basing a purchase on estimates alone. That posture is rarer than it should be.
The reservations concern governance rather than craft. There is no privacy policy, no terms and no legal notice; the operator is anonymous, described only as a small group of enthusiasts; and the domain was registered in July 2026, so nothing here has a track record yet. The service is entirely free with no visible business model, which is pleasant today and says nothing about next year. It is English-only, offers no API, and remembers nothing between visits.
Treat it for what it is: a well-built, transparent decision aid for anyone curious about running AI offline on hardware they already own, and a quick sanity check before paying for a larger memory tier. Use its verdicts to narrow the field and its glossary to understand them, confirm anything that carries real consequences, and never mistake a careful estimate for a measurement.
- Choosing a selection results in a full page refresh.
- Opens in a new window.