CanIRun.ai
CanIRun.ai detects your GPU or Apple chip in the browser and grades 104 open-weight models, from chat and coding to image and video, by whether they actually fit, with VRAM requirements and speed estimates. Free, no account required.
What is CanIRun.ai?
CanIRun.ai is a free public site built around a single question: can this machine run an open-weight AI model, and how well? It ignores hosted assistants entirely and covers only models you can download and execute yourself.
Everything starts in the browser. Three APIs identify the hardware: WebGL, through the WEBGL_debug_renderer_info extension, returns the GPU name and vendor; WebGPU exposes the adapter's device and architecture; and the Navigator object supplies the core count and approximate RAM, completed by a short CPU benchmark of roughly 30 milliseconds. The result is matched against a built-in database of around forty NVIDIA, AMD and Intel GPUs and about a dozen Apple Silicon chips, each carrying its memory capacity and bandwidth.
From there the site computes, for every model, the VRAM required at each quantization level: parameters multiplied by bits per weight, divided by eight, plus a fixed 0.5 GB runtime overhead and a ten percent margin for the KV cache. A 70B model at Q4_K_M therefore lands near 39 GB. Each model then receives a score out of 100 built from estimated speed (55%), memory headroom (35%) and a capped quality bonus, with a 0.65 penalty when the fit is tight. Scores map to six grades, from Runs great down to Too heavy.
The catalogue holds 104 models from 22 labs, among them Alibaba, Meta, Google, DeepSeek, OpenAI, Mistral AI, Microsoft and Tencent, split between chat and coding, image, video and lightweight models, alongside 284 device pages covering Apple Silicon, NVIDIA, AMD, Intel, phone GPUs and single-board computers. Around this core sit a two-device comparator, an exportable tier list, a WebGPU playground that runs models directly in the browser, a glossary and a page detailing the whole methodology with its formulas.
The last step is execution: each model page carries a runai command, the companion command-line tool installed with a single curl line on macOS, Linux or WSL, which picks a quantization, downloads the GGUF weights and opens a local chat through llama.cpp. Ollama and LM Studio work with the same files. The site is built with Astro by midudev, and its source code is public on GitHub.
What it does
- Detects the visitor's GPU, VRAM, memory bandwidth, RAM and CPU cores from the browser
- Grades 104 open-weight models from S to F for the machine it has detected
- Estimates tokens per second and the VRAM footprint of every quantization level
- Opens a dedicated page for any of 284 devices and for every model in the catalogue
- Compares two devices side by side, model by model, with a win count
- Runs a model in the browser through WebGPU, with nothing to install
- Hands over the runai command that downloads and launches the model locally
When to use CanIRun.ai / When not to
A quick filter to help you decide if CanIRun.ai is the right fit.
When to use CanIRun.ai
- Developers who want to know whether their existing GPU or Mac can run an open-weight model before downloading several gigabytes of weights
- Buyers sizing a workstation or a laptop, using the 284 device pages and the two-device comparator to choose between VRAM tiers
- Machine learning and MLOps engineers who need a quick VRAM and tokens-per-second sanity check for each quantization level
- Designers, video editors and motion artists hunting for image or video models that fit a consumer graphics card
- Students and lecturers learning the vocabulary of local inference: parameters, quantization, GGUF, mixture of experts, memory bandwidth
When not to use CanIRun.ai
- Teams that need a hosted model or a cloud API, since the site covers only open-weight models you download and run yourself
- Anyone who needs measured benchmarks, because every figure here is an estimate derived from browser APIs and vendor spec sheets
- Organisations with procurement or compliance requirements, as the domain publishes no terms of use, privacy policy or legal notice
- Users looking for a mobile application or a documented public API, neither of which exists
- Non-English speakers, since the interface and all the documentation are English only
How to use CanIRun.ai
A typical end-to-end flow, from setup to results.
- Open canirun.ai in a recent browser and let the page read your GPU and memory
- If detection misses your chip, pick it by hand from the 284 device pages
- Read the ranking of open models that fit, sorted by grade from S down to F
- Filter by task, whether chat and coding, image, video or lightweight, and by lab, licence or commercial use
- Open a model page and read its quantization table, with the VRAM and quality of Q4_K_M, Q6_K, Q8_0 and the rest
- Check the licence badge before committing, since several catalogued models are non-commercial
- Copy the runai command shown on the model page, for instance runai run qwen3.5-4b
- Install runai with the single curl command on macOS, Linux or WSL, reviewing the script first as the site itself suggests
- Or skip the install entirely and try the model in the WebGPU playground, adjusting system prompt, temperature, max tokens and top-p
- Use the compare page or the exportable tier list to weigh two machines against each other
Pros & Cons
Pros
- Entirely free, with no account, no sign-up wall and no paid tier
- The scoring methodology is published in full, formulas and thresholds included
- Detection and scoring run client-side, and the site states that nothing is sent to a server
- The source code of the site is public on GitHub
- Broad and current catalogue: 104 models, 284 devices, 22 labs
- The path is complete, from diagnosis through to a working local model via runai, Ollama or LM Studio
- Licences are shown per model, with commercial and non-commercial use clearly separated
Cons
- Every figure is an estimate rather than a measurement, and the site says so in capitals
- Browser detection is imprecise: vague GPU names, approximate RAM, bandwidth taken from spec sheets
- Thermal throttling, driver versions, background processes and operating system memory pressure cannot be measured and are not accounted for
- The built-in GPU database described on the methodology page covers far fewer chips than the 284 device pages suggest
- No terms of use, privacy policy or legal notice is published on the domain
- No support contact: the only email address on the site is reserved for advertising campaigns
- English only, with no mobile application and no documented public API
Pricing & Plans
The tool is free of charge in its entirety: there is no paid plan, no pricing page and no payment of any kind is required to use the site, the WebGPU playground or the runai command-line tool, so no entry price applies. The only paid item on the domain is a display advertising slot, which concerns advertisers rather than people using the tool.
- full access to the model rankings
- the device pages
- the two-device comparator
- the tier list
- the WebGPU playground and the runai command-line tool
- with no account and no time limit
Data, GDPR & hosting
A consolidated view of how CanIRun.ai handles your data.
GDPR overview
There is no mention of the GDPR anywhere on canirun.ai. The domain publishes no privacy policy, no cookie policy, no consent banner, no data protection officer and no Article 27 representative; targeted probes on /privacy, /terms and /legal all return 404, and the complete sitemap contains no legal page. The only privacy statement is technical: the methodology page asserts that detection and scoring happen entirely in the browser and that no data is sent to any server, which would leave little to regulate. The robots.txt file cites Article 4 of EU Directive 2019/790, but that concerns copyright reservations, not data protection. For a project whose author appears to be established in the European Union, this complete absence of published legal documentation is worth noting.
Who owns the data?
No terms of use or privacy policy is published on the domain, so nothing contractually assigns ownership of anything. The only statement on the subject is the Privacy and transparency section of the methodology page, which says that hardware detection and scoring run entirely in the browser and that the GPU name, RAM and benchmark results never leave the device. In the playground, downloaded models are cached in the visitor's browser and conversations are stored locally, persisting across sessions, which leaves them under the visitor's control. In practice you keep everything, but by the absence of any published agreement rather than by an explicit commitment.
Reuse rights
The site publishes no terms governing reuse, so no permission mechanism is documented in either direction. What it does state is that the visitor's hardware data serves only the local calculation of scores and rankings and is never sent to a server. The third-party figures displayed, namely download and like counts from the HuggingFace API and pull counts and tags from the Ollama library, are fetched by the site at build time rather than from the visitor. The robots.txt file signals search=yes, ai-train=no and use=reference for the site's own content, and blocks GPTBot, ClaudeBot, CCBot, Google-Extended and several other crawlers.
Data retention & training
Hosting summary
No hosting jurisdiction is declared for user data, for the simple reason that the site declares no user data collection at all. The methodology page states that hardware detection and scoring run entirely in the browser and that nothing is sent to a server. In the playground, downloaded models are cached in the visitor's browser and conversations are stored locally, so whatever persists stays on the visitor's own machine. As for the site itself, the domain resolves to 172.67.201.155, an anycast address belonging to AS13335, Cloudflare Inc., with a node geolocated in San Francisco, United States; that describes how the pages are delivered, not where any personal data lives. The model statistics on display are fetched by the site from the HuggingFace API and the Ollama library at build time, again without involving the visitor. Because the domain carries neither a privacy policy nor terms of use, none of this is contractually documented: it rests on a technical statement on the methodology page and on what the pages themselves do.
Things to keep in mind
Risks and trade-offs to weigh before adopting CanIRun.ai.
- The grades are estimates, so buying a graphics card on the strength of a browser-detected score can turn into an expensive mistake
- Installing runai means piping a remote script into bash; the site itself suggests reading the installer at canirun.ai/runai/install.sh first
- Several catalogued models carry non-commercial or conditional community licences, and using one in a paid project without checking is a genuine legal risk
- Running a model locally removes the provider's safety layer: the output is unfiltered and entirely your responsibility
- With no terms, no privacy policy and no support contact, there is nothing to rely on if something goes wrong
- Mixture-of-experts models load every expert into memory, so the active parameter count understates what a machine actually needs
- Downloading multi-gigabyte weights on a metered or shared connection carries a cost the site does not show
Setup & Integrations
Technical difficulty
Using the site itself requires nothing: open it in a recent browser and the ranking appears, with no account and no install. The playground is barely harder, needing a WebGPU-capable browser and the patience to download the chosen model once. The runai command-line tool is the only step that assumes some comfort with a terminal: a single curl command on macOS, Linux or WSL, then runai run followed by a model identifier. The real difficulty lies elsewhere, in owning suitable hardware and downloading multi-gigabyte weights, not in the tool.
Deployment
Integrations
Supported languages
Behind CanIRun.ai
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
How much VRAM do I need to run a local AI model?
Does CanIRun.ai cost anything?
Is my hardware data sent anywhere?
Can I try a model without installing anything?
How do I actually run a model the site recommends?
What is quantization and why does it matter here?
Can I run local models on a Mac without a dedicated GPU?
How reliable are the grades and the speed estimates?
Does the site publish terms of use or a privacy policy?
Who is behind CanIRun.ai?
Should you pick CanIRun.ai?
CanIRun.ai does one thing, and does it with unusual transparency. It answers the question every developer curious about local AI eventually asks, namely whether a given model will run on a given machine, and it answers in seconds without an account, a paywall or a server round trip. The methodology page is the strongest argument in its favour: the detection APIs, the VRAM formula, the score weighting and the grade thresholds are all published, with the code that produces them, and the source of the site itself is on GitHub. Few directories expose their own arithmetic that plainly.
The caveats are just as clear, and the site states them first. Every number is an estimate built on browser APIs and vendor spec sheets rather than on measurement; thermal throttling, driver versions and background load simply cannot be seen from a web page. The hardware database described in the methodology is narrower than the 284 device pages suggest. And the project is very young: the domain was registered on 9 March 2026, the first archive capture dates from the same day, and it is maintained by a single developer.
The more serious gap for professional use is legal. There are no terms of use, no privacy policy, no legal notice and no support address; the only email on the site is reserved for advertising. That matters less for a tool that claims to send nothing to a server, but it leaves an organisation with nothing to point at.
Taken for what it is, a free and well-documented diagnostic and catalogue for open-weight models with a working path through to runai, Ollama or LM Studio, CanIRun.ai is genuinely useful. Treat its grades as guidance for a hardware decision, and verify anything that carries a budget.
- Choosing a selection results in a full page refresh.
- Opens in a new window.