RTX laptop
An AI-optimized Windows/Linux laptop with a discrete NVIDIA GPU (RTX 4080/4090 Laptop, 12–16GB VRAM). Less memory than an Apple Silicon machine, but CUDA gives it the fastest prompt-processing in this tier — the right pick when executor latency matters more than model size.
What fits
| VRAM | Good picks | Tier |
|---|---|---|
| 12–16GB | Qwen3 14B AWQ, Qwen3 8B, Gemma 3 12B | swarm / executor |
VRAM is the hard limit here — the model plus its KV cache must fit the card, with no unified-memory fallback. python scripts/fit_device.py --memory 16 --backend cuda picks a model that fits and prints the launch command plus endpoints.yaml stanza.
Serving
vLLM (CUDA), via hardware/personal-devices/configs/serve-cuda.sh — AWQ (4-bit) fits the most model per GB of VRAM:
MODEL=Qwen/Qwen3-14B-AWQ ./configs/serve-cuda.sh
Register under tier executor (or swarm for the smaller models) in scripts/endpoints.yaml.
Role
The fastest executor of the consumer options for latency-sensitive work. Same caveat as any laptop: don't rely on it as your only executor if it also needs to survive being unplugged or closed mid-run — prefer a desktop tower for unattended serving.