CanIRun.ai
canirun.ai- Category
- AI Tools
- Rank
- No. 2822Tools index
Previous survey · No. 2747 ·
- Pricing
- Free
- Type
- APP
- Use case
- Models: Train & Run
- Interfaces
- Web
- Builder
- @CamilleRoux
- GitHub
- 550 stars
- Date
About
A browser-based hardware compatibility checker for AI models. It detects your GPU, CPU, and RAM specifications to tell you which AI models you can run locally and how well they'll perform.
What it does
A website that answers one question before you download anything: will this open-weight model run on my machine, and how fast? The page reads your graphics chip name, memory and a quick processor timing test straight from the browser, then matches the chip against a built-in table of memory size and bandwidth figures. For each model it works out the memory needed at seven compression levels, estimates tokens per second from memory bandwidth, and assigns a letter from S down to F. The README says nothing is sent to a server. The same engine is also exposed as a small JSON API.
Why it's ranked here
Picking a local model usually means guessing at memory math. This project does that math openly and shows its inputs. The hardware table covers NVIDIA consumer, laptop, workstation and datacenter cards, AMD Radeon, Intel Arc, Apple Silicon and mobile chips, with a source note on each block. A test suite pins how real driver strings map to cards, including cases where one model number could be mistaken for another. It is MIT licensed, and a companion command line tool turns the recommendation into a download and a local server. The limit: every number is an estimate from specifications, not a measured run.
What's good
The API is practical. Pass only a graphics card name and it fills in memory and bandwidth from its table; omit the card and it treats the machine as processor only. Apple chips are handled as shared memory rather than dedicated video memory. When you skip the compression level, it picks the highest-quality one that still fits comfortably. Answers come back with plain-English notes, such as how much memory headroom remains for context. Model pages include one-click install commands for Ollama and LM Studio. Keyboard shortcuts and filters by use case, provider, dense versus mixture-of-experts, and tool use make the catalogue quick to scan.
Tradeoffs
Speed figures are approximations derived from memory bandwidth, and memory needs come from parameter count times a fixed bytes-per-weight table plus a flat overhead factor, so context length and runtime differences are not modelled. Browser detection depends on the graphics name the browser exposes; an unknown chip returns no match. With no bandwidth given, the API assumes dual-channel DDR5 at 50 GB/s. The companion runtime needs Bun and pnpm, and its README says recommendation quality is currently tuned for Apple Silicon Macs. That tool also sends anonymous telemetry unless you set an opt-out variable.
How to use it well
Use the site as a first filter before downloading tens of gigabytes: check which compression level fits, then confirm with a real run. Builders of PC configurators or dashboards can call the recommend endpoint with a hardware profile and a use case to get a ranked list. If you want to go straight from recommendation to chatting, the companion tool can pull a model, open a terminal chat, serve an OpenAI-compatible local API, and benchmark tokens per second on your actual hardware. The site itself runs no model, so treat its grade as a shortlist, not a measurement.
Technical notes+
The site is Astro with two workspace packages: packages/compatibility/src/index.ts holds the GPU database (vram, bw, cores per card) and scoring, and src/lib/hardware.ts and src/data/models.ts simply re-export the packages. src/lib/compatibility-api.ts maps the public hardware input onto the internal HardwareInfo shape, enriches from matchGPU or matchApple, falls back to 50 GB/s for CPU-only profiles, and picks the best-fit quant by sorting quants by bits. scripts/scrape-models.ts pulls Hugging Face metadata, uses a bytes-per-parameter table with a 1.2 runtime overhead, and estimates MoE active parameters. tests/hardware.test.ts covers renderer-string matching. packages/runai/src/cli.ts is a Bun CLI built on node-llama-cpp. AGENTS.md tells agents to skip typechecks and tests unless asked.
Observed
- License
- MIT
- Language
- TypeScript
- Framework
- Astro static site with Tailwind CSS
- Interfaces
- Web app, JSON API (models, compatibility, recommend), companion CLI
- Package manager
- pnpm workspace; companion CLI requires Bun
- Hardware detection
- Client-side via WebGL, WebGPU and a CPU micro-benchmark
- Quantization levels
- Seven, from Q2_K to F16
- Companion runtime API
- OpenAI-compatible local endpoints
- Tests
- Vitest suite for hardware matching and categorisation
Read from README.md, package.json, AGENTS.md, packages/runai/README.md, packages/runai/package.json, packages/runai/src/cli.ts, packages/compatibility/src/index.ts, src/lib/hardware.ts, src/data/models.ts, src/lib/compatibility-api.ts, tests/hardware.test.ts, scripts/scrape-models.ts.
What it can do
Detect hardware specifications
User's computer system → GPU, CPU, and RAM specifications
Check AI model compatibility
Hardware specifications and AI model requirements → Compatibility status for specific AI models
Predict AI model performance
Hardware specifications and AI model → Performance metrics and expected execution speed
Recommend runnable AI models
Detected hardware capabilities → List of AI models that can run locally
Analyze system requirements
Specific AI model → Minimum and recommended hardware requirements
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.