# modelread > Hardware-aware fit checker for Hugging Face models: given a repo and a device (GPU, Mac, phone, multi-GPU node or > cloud instance) it returns whether it fits, the best quant, decode/prefill tokens per second as a calibrated range, > the longest context that fits, memory placement (VRAM/RAM/SSD), and the exact llama.cpp / vLLM / MLX launch command. > Memory comes from the repository's real files and tensor headers, not parameter counts. Use it before downloading, quantizing or serving a model; when a user asks "will X run on my Y", "which quant fits", "how many tok/s", "what's the best model for task Z on my machine", or needs a launch command / serving config. ## MCP (preferred for agents) - Remote MCP server, streamable HTTP, no auth: https://modelread.com/mcp - Claude Code: `claude mcp add --transport http --scope user modelread https://modelread.com/mcp` - Codex: `codex mcp add modelread --url https://modelread.com/mcp` ยท Gemini CLI: `gemini mcp add --transport http modelread https://modelread.com/mcp` - Tools: estimate_fit(repo, hardware, ctx), compare_hardware(repo, hardware[]), find_models(task, hardware, ...), search_models(query), list_hardware(filter), ask(question). All read-only and deterministic. - Hardware strings: preset ids (rtx4090, m4max-128, 8x-h100-sxm, iphone-17pro-12, aws-p5.48xlarge), "2x-rtx3090", free text ("MacBook Pro M3 Max 128GB"), or custom "gpu=24,ram=64,bw=936". ## HTTP API - [Estimate](https://modelread.com/api/estimate?repo=Qwen/Qwen3.6-27B&hw=rtx4090&ctx=32768): GET repo, hw, ctx, kv (f16|q8_0|q4_0), ram - [Ask (SSE)](https://modelread.com/api/ask/stream): POST {"q": "...", "hw": ""}; events status, query, draft, final - [Hardware catalog](https://modelread.com/api/hardware): every device preset with memory, bandwidth, compute, backend, $/hour - [What runs where](https://modelread.com/api/matrix?hw=rtx4090): nightly best variant per popular model per reference device - [Search](https://modelread.com/api/search?task=coding&q=qwen): index search - [OpenAPI](https://modelread.com/openapi.json) ## Reading the output - verdict: fits | fits-tight | offload (works, slower: part in system RAM) | no | unknown - decode_tps {lo, mid, hi, calibrated, basis}: quote the range, not just mid; "calibrated": fitted on measured devices - max_ctx_fit: longest context that fits with the chosen KV dtype - launch: copy-paste command; for multi-GPU nodes a tensor-parallel vLLM line - reasons / unknowns / warnings: say these to the user when present ## Pages - [Ask](https://modelread.com/): natural-language questions - [Model pages](https://modelread.com/m/Qwen/Qwen3.6-27B): /m/{org}/{repo}, fit on your device and across device classes - [Device pages](https://modelread.com/hw/rtx4090): /hw/{id}, best models for one device - [Methodology](https://modelread.com/methodology): how memory and speed are computed and calibrated - [Docs](https://modelread.com/docs): quickstart, MCP tool reference, HTTP API, hardware strings, reading results, limits ## Optional - [Agent skill](https://modelread.com/skill.md): SKILL.md for Claude Code and other skill-aware agents - [Server card](https://modelread.com/.well-known/mcp/server-card.json)