Will it run on your hardware?
Ask in plain words or paste any Hugging Face model. modelread reads the real file sizes and tensor headers, picks the quant that fits, and tells you tokens/sec as an honest range, the longest context, and the exact command to run it.
Speed as a range
Every tok/s is a band with its source: measured on that device, calibrated, or an uncalibrated estimate.
Phones to 8× H100
GPUs from every vendor, Apple Silicon, Jetson, phones, Raspberry Pi, VPS and cloud instances with $/hour.
Real files, not guesses
Sizes come from the repo's actual weights, KV cache from the architecture, so MoE, SWA and hybrid models are right.
Built for agents
The same answers over MCP and JSON, so Claude Code, Cursor or Codex can size and launch models for you.
Popular right now
Open a model to see how it runs on your device, and on every other.
The model index is warming up.
Let your coding agent ask
One line adds modelread to Claude Code. Cursor, VS Code, Codex, Gemini CLI, Zed and LM Studio are on the Connect page.
claude mcp add --transport http modelread https://modelread.com/mcp- estimate_fit: will repo X run on device Y, and how
- find_models: best model for a task on a device
- compare_hardware: one model across up to 8 devices