我的電脳跑得動 AI 嗎？

偵測你的 GPU 或 Mac，把真的跑得動的開源模型排給你看。

GPUVRAMBWRAMCores—

最適合以下装置的開源模型： 你的装置

跑得動写程式的模型嗎？

対話、程式与推理

跑得動図像模型嗎？

在本機生成図像

跑得動影片模型嗎？

在本機生成影片

跑得動超小模型嗎？

装置端，4B 以下

104 models · 292 devices · 22 labs

Run open AI models on your own hardware

We detect your GPU or unified memory and rank open models you can run locally — chat, coding, image and video. A 9B model fits an RTX 4060; an M4 Max can take larger ones. See the scoring method, docs or install runai.

  1. 01

    Open CanIRun.ai

    Visit the homepage in a desktop browser. No account or extension is required.

  2. 02

    Detect or pick your hardware

    Let the page read your GPU and memory, or browse the device list if detection misses your chip.

  3. 03

    Compare recommended models

    Review the top open models for text, coding, image and video that fit the detected VRAM or unified memory.

  4. 04

    Run a model locally

    Open a model page and launch it with runai, Ollama or LM Studio using the suggested quantization.

How much VRAM do you need?

Figures assume Q4_K_M plus a small runtime overhead. Mixture-of-experts models still load all experts into memory, even when only a few are active per token.

Approximate local AI model sizes by available GPU memory
MemoryWhat usually fitsExample device
8 GB3B–9B chat models at Q4, plus tiny image models.RTX 4060
12 GB12B–14B dense models, or smaller mixture-of-experts.RTX 4070
16 GB20B–27B at Q4, comfortable 14B at higher quality.Apple M4
24 GB30B dense models, mid-size MoE, or local video.RTX 4090
32 GB+Larger open MoE and high-quality image generation.RTX 5090

Popular devices

All 292 devices

Labs behind the models

Open models worth running locally

A short mix of chat, image and video models. The list above still ranks whatever fits the hardware detected on this page.

Guides

If you want the vocabulary or the scoring math, these two pages cover it.

Common questions about local AI

How much VRAM do I need to run a local AI model?

A 7B–9B chat model usually fits in 8 GB of VRAM at Q4 quantization. 12–16 GB covers most 12B–27B models. 24 GB and up opens 30B dense models, mid-size mixture-of-experts, and local image or video generation. On Apple Silicon the limit is unified memory, not a separate GPU frame buffer.

Can I run local AI models on a Mac without a dedicated GPU?

Yes. Apple Silicon Macs share RAM between CPU and GPU, so an M4, M4 Pro or M4 Max can run open models that would otherwise need a discrete graphics card. Bandwidth still matters: a Max chip generates tokens faster than a base M-series chip with the same memory.

M4 modelsM4 ProM4 Max

What is the difference between running AI locally and using ChatGPT?

Local models run on your machine. Prompts never leave the device, there is no subscription meter, and you can keep working offline. Cloud assistants are usually stronger on the hardest tasks, but they need a network and send your data to a provider. CanIRun.ai is for the open-weight models you can download and run yourself.

What is quantization and why does it matter?

Quantization stores model weights in fewer bits so the file is smaller and uses less memory. Q4_K_M is the usual sweet spot for local chat: much smaller than full precision, with only a modest quality drop. Higher formats such as Q8 or F16 look closer to the original model but need more VRAM.

Quantization guide

How does CanIRun.ai know which models fit my hardware?

The site reads GPU, RAM and CPU hints from the browser, matches them against a hardware database, then scores each open model on estimated speed, memory headroom and quality. You can also skip detection and open a device page for a specific GPU, Apple chip, phone or Raspberry Pi.

Scoring methodologyBrowse devices

How do I download and run a recommended model?

Open any model page and copy its runai command, or use Ollama or LM Studio with the same GGUF weights. runai picks a quantization for your machine, downloads the file, and starts a local chat with llama.cpp.

Install runai

Want the scoring formulas, or a side-by-side of two checkpoints? Read how detection works, open the model compare tool, or check the tier list.

由 midudev 為本機 AI 社群製作GitHub

数值由瀏覧器 API 估算，実際規格可能不同。WebGPU資料来源： llama.cpp, Ollama 与 LM Studio.

所有產品名称、標誌与品牌均為其各自所有者的財產。Apple、NVIDIA、AMD、Intel、Qualcomm 以及本站提及的所有 AI 模型名称均為其各自持有者的商標或註冊商標。本站与上述任何公司均無関聯，亦未獲其背書。