Detect your GPU or Mac and see which open LLMs, coding, image and video models you can run locally. VRAM requirements, speed estimates and grades.

Ccanirun.aiᐧby midudev8 min read

Why

Cloud AI APIs are expensive, rate-limited, and send your data to third parties. Running models locally gives you privacy, speed, and zero cost per token — but only if your hardware is up to the job.

CanIRun.ai answers that question instantly. Open the site, let it detect your hardware, and get a curated set of best-pick recommendations grouped by use case (general, coding, reasoning, vision and lightweight) — the top open models that actually run well on your device. Prefer to explore? Switch to Browse all for the full compatibility report across 55+ open-weight models with grades from S to F.

How It Works

Browser APIs → Hardware Detection → Per-Use-Case Ranking → Best-Pick Recommendations

1. Hardware detection runs entirely client-side using WebGL, WebGPU, navigator.deviceMemory and a lightweight CPU micro-benchmark.
2. Each model's VRAM requirements are calculated across 7 quantization levels (Q2_K → F16) from parameter count.
3. A scoring algorithm combines run status, estimated tokens/second, memory headroom and model size into a letter grade (S–F).
4. Results are displayed instantly — nothing is sent to any server.

Supported hardware

Features

• Best-pick recommendations — the top open models for your device, grouped by use case (general, coding, reasoning, vision, lightweight), quality-ranked but gated by what actually runs well
• Zero-install hardware detection — CPU cores, RAM, GPU model, VRAM and memory bandwidth identified from the browser
• 55+ curated open models — from Qwen 3 0.6B up to GLM-5.2 753B and Kimi K2.6 1T, pruned to the ones worth running (no stale duplicates)
• 7 quantization levels per model — Q2_K, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16 with computed VRAM sizes
• S–F grading system — instant letter grade based on your hardware vs. model requirements
• Tokens/second estimates — approximate inference speed from memory bandwidth data
• Filters — by use case (chat, code, reasoning, vision), provider, architecture (dense / MoE), features (tool use, thinking)
• Search & keyboard shortcuts — / to search, j/k to navigate, Enter to open, v to switch view
• Three view modes — compact grid, detailed grid, and list
• Tier list — shareable S–F tier list you can export as an image
• Model detail pages — per-quant compatibility table, one-click Ollama / LM Studio / llama.cpp install commands
• OG images — dynamically generated social preview images for every model
• SEO — Schema.org structured data, sitemap, semantic HTML
• View Transitions — smooth page animations via Astro Client Router

Model Catalog

Models from Meta, Google, Alibaba, DeepSeek, Mistral AI, Microsoft, NVIDIA, Liquid AI, Z.ai, Moonshot AI, OpenAI and the community:

API

The same compatibility engine that powers the site is exposed as a small JSON API, so you can integrate CanIRun.ai into dashboards, PC configurators, CLI tools or custom assistants. All endpoints are CORS-enabled and return application/json.

Hardware profile — all POST endpoints accept a hardware object. Provide a GPU name and we enrich VRAM/bandwidth from the internal database, or pass explicit values:

{
  "hardware": {
    "cpu": { "name": "AMD Ryzen 7 5800X", "cores": 8, "threads": 16 },
    "ramGb": 32,
    "gpu": { "name": "NVIDIA RTX 3060", "vramGb": 12, "memoryBandwidthGbps": 360 }
  }
}

Apple Silicon ("gpu": { "name": "Apple M3 Max" }) is detected automatically and treated as unified memory. Omit gpu for a CPU / integrated-GPU profile.

POST /api/compatibility

curl -X POST https://canirun.ai/api/compatibility \
  -H 'content-type: application/json' \
  -d '{ "hardware": { "ramGb": 32, "gpu": { "name": "NVIDIA RTX 3060" } },
        "modelId": "llama-3.1-8b", "quantization": "Q4_K_M" }'

{
  "compatible": true,
  "status": "comfortable",          // comfortable | tight | cpu-offload | insufficient | unknown
  "grade": "A",                     // S–F
  "score": 82,
  "modelId": "llama3.1-8b",
  "quantization": "Q4_K_M",
  "recommendedQuantization": "Q8_0",
  "estimated": {
    "tokensPerSecond": 55,
    "modelSizeGb": 3.9,
    "vramRequiredGb": 4.6,
    "ramRequiredGb": 7.5,
    "memoryHeadroomGb": 7.4
  },
  "notes": ["The model should fit comfortably in GPU memory.", "..."]
}

quantization is optional — when omitted, the best-fitting quant is used.

POST /api/recommend

curl -X POST https://canirun.ai/api/recommend \
  -H 'content-type: application/json' \
  -d '{ "hardware": { "ramGb": 64, "gpu": { "name": "RTX 4090", "vramGb": 24 } },
        "useCase": "code", "limit": 5 }'

Returns a ranked recommendations array (each with quant, grade, status and estimated tokens/second). useCase and limit (1–25, default 5) are optional.

Tech Stack

Getting Started

Prerequisites: Node.js 18+ and pnpm

# Clone the repo
git clone https://github.com/midudev/canirun.ai.git
cd canirun.ai

# Install dependencies
pnpm install

# Start dev server
pnpm dev

Open localhost:4321 to see the site.

Commands

Project Structure

packages/
├── models/
│   └── src/index.ts        # 90+ AI model definitions with quant calculations (edit here)
├── compatibility/
│   └── src/index.ts        # GPU/Apple/Mobile DBs + compatibility scoring (edit here)
└── runai/                   # CLI for running models locally

src/
├── data/
│   ├── models.ts            # Re-exports @canirun/models (built from packages/models)
│   └── hf-stats.json        # HuggingFace download/like counts
├── lib/
│   ├── hardware.ts         # Re-exports @canirun/compatibility
│   └── og.ts               # OG image generation utilities
├── pages/
│   ├── index.astro         # Home — model grid with filters & search
│   ├── tier.astro          # Tier list — S–F ranking with image export
│   ├── model/[id].astro    # Model detail — quants, compatibility, install
│   ├── api/                # JSON API — models, compatibility, recommend
│   └── og/                 # Dynamic OG image endpoints
├── components/
│   └── NavHeader.astro     # Site navigation
├── layouts/
│   └── Layout.astro        # Base layout with SEO, fonts, transitions
├── icons/                  # SVG icon components
└── styles/
    └── global.css          # Theme tokens, Geist fonts, dark mode

Contributing

Contributions are welcome! Some ways to help:

• Add a model — add an entry to the STATIC_MODELS array in packages/models/src/index.ts following the existing pattern (the AIModel interface at the top of that file).
• Improve hardware detection — extend the GPU/Apple/Mobile databases in packages/compatibility/src/index.ts
• Report inaccurate results — open an issue with your hardware info and the model in question
• Fix bugs or improve UI — PRs are appreciated

Author

Created by midudev · @midudev