Skip to content
modelread

Loading devices…

Remembered on this device. Multi-GPU: search “2x”, “8x”, “node”.

Hardware strings

Every hardware / hw argument accepts four forms.

Preset ids

330+ presets with memory, bandwidth, compute and backend. Browse them on Devices or call list_hardware. Examples: rtx4090, rtx3060-12, rx7900xtx, arc-b580,m4max-128, m3ultra-512, dgx-spark, strix-halo-128, jetson-thor,h100-sxm, 8x-h100-sxm, iphone-17pro-12, pixel-10pro-16, rpi5-16.

Multi-GPU

<N>x-<gpu id>, e.g. 2x-rtx3090, 4x-rtx3090. Multi-GPU nodes also get a tensor-parallel (vLLM / SGLang) estimate.

Free text

“two RTX 3090s”, “MacBook Pro M3 Max 128GB”, “Galaxy S25 Ultra”, “AWS p5.48xlarge”, “DigitalOcean H100 droplet”. Matched to the closest preset; the response names the device it used.

Custom spec

  • Discrete GPU: gpu=24,ram=64,bw=936 (VRAM GiB, system RAM GiB, memory bandwidth GB/s; optional ram_bw, backend)
  • Unified memory: unified=1,ram=96,bw=256

Override just the system RAM of a preset with ram (API) or ram_gb (MCP).

Askwhat should I run?
Browsebest models per device
Devicesall hardware
Connect your agentMCP · API
How the numbers workcalibration