Hardware
Will it run on my laptop?
CLAIR runs its AI model on your own machine. Quality scales with RAM: a bigger machine unlocks a stronger model, but the base tier runs on effectively any Windows laptop.
Quality tiers
Five tiers, gated on RAM alone
Tier availability is a flat RAM check, not tied to whether you have a graphics card. A machine with no GPU can still select any tier its RAM allows: acceleration only changes how fast a tier answers, never whether it unlocks. Each tier has exactly one fixed default model, so the numbers below are stable, not the fastest model that happened to be available that week.
| Tier | RAM needed | Default model | Model size | Context window | Prompt column limit |
|---|---|---|---|---|---|
| 1. Lightning | Any machine | Qwen 3.5 2B | 2.7 GB | 4,096 | 50 |
| 2. Fast | 8 GB | Qwen 3.5 4B | 3.4 GB | 32,768 | 100 |
| 3. Balanced | 12 GB | Qwen 3.5 9B | 6.6 GB | 32,768 | 200 |
| 4. Powerful | 16 GB | Gemma 4 12B | 7.6 GB | 32,768 | 250 |
| 5. Max | 32 GB | Qwen 3.6 27B | 17.0 GB | 32,768 | 400 |
RAM figures above are the numbers CLAIR shows you. The internal check sits a hair below each one, because Windows reports total physical RAM somewhat under the number on the box: a laptop sold as "16 GB" typically detects around 15.3 GB, so the gate is set to still clear on that exact machine.
Acceleration
Graphics cards CLAIR detects
CLAIR probes for a supported GPU at startup and configures Ollama to use it automatically. There is nothing to install or configure yourself.
NVIDIA
Detected via CUDA. CLAIR sizes the model against your card's VRAM, with a small margin for quantization, and falls back to sizing off RAM if a model doesn't fit in VRAM but the machine has room to spare.
Intel Arc (dGPU and iGPU)
Driven through Ollama's own Vulkan backend, not a separate build. On a Core Ultra laptop with an Arc iGPU, CLAIR sets OLLAMA_IGPU_ENABLE=1 so the integrated GPU is actually used instead of silently falling back to CPU. Measured on an Arc 140T: the configured model loaded 100% into VRAM.
AMD
Detected when PyTorch was built with ROCm/HIP support. CLAIR reads the real card name and VRAM through the same CUDA-shaped API ROCm exposes, and sizes the model off VRAM the same way the NVIDIA path does.
Apple Silicon
Detected as a Metal backend, sized against unified memory (roughly three quarters of system RAM budgeted for the model). This detection exists in CLAIR's hardware code today; a macOS installer has not shipped yet, see the download page for the current, honest status.
Disk
What it needs on disk
The installer itself is roughly a gigabyte. The first time you launch CLAIR, it downloads two more things once and reuses them after that: a library pack of data-science tools (needed for Predict/Train, NLP, forecasting and advanced statistics) and the AI model your hardware tier qualifies for, ranging from about 2.7 GB on the smallest tier to 17 GB on the largest. After that first run, CLAIR works fully offline.
Summary
Minimum and recommended
Minimum
Windows 10 or 11, any amount of RAM. The Lightning tier has no RAM floor and runs on a CPU-only machine, just more slowly than an accelerated one.
Recommended
16 GB of RAM gets you the Powerful tier and a genuinely capable model. 32 GB unlocks the Max tier. A GPU is not required, but it makes every tier answer faster.
What gets slower on a small machine
On CPU-only hardware, larger tiers get noticeably slower to answer: measured full-answer times on the same CPU-only tier ladder ran about 18 seconds on Lightning, 25 on Fast, 50 on Balanced, 75 on Powerful, and 150 on Max. Because of this, CLAIR's own recommendation logic never defaults you to a tier whose typical answer would take more than 30 seconds, even if your RAM would technically allow it: on a CPU-only machine the recommended default rarely rises past Fast, regardless of how much RAM you have. Every tier your RAM allows stays selectable if you want the extra quality and are willing to wait. If one answer does take a while, CLAIR notices: any single answer over 20 seconds offers a one-click switch down a tier, naming the real measured time and the model it would switch to.