# Nodegrove

> Will it run? Where? For how much? Independent answers for running open AI models: whether a model or ComfyUI workflow fits a GPU, what it costs and how fast it runs, and where to run it. Free tools, open data and an MCP server.

Every number on the site is one of four kinds, and the page says which: measured (we rent the card and run the job), computed (from the model's config.json and file sizes, with the formula shown), read (a provider's own price or limit, with the date it was read), or typed by the user. The site earns from referral links; a referral link is labelled as one next to it and carries rel="sponsored", and a payment never changes a verdict or a ranking. Details: https://nodegrove.io/independence

## Pages

- [Home](https://nodegrove.io/): the free tools, where to run a model or workflow (your own GPU, Comfy Cloud, a GPU by the hour, or pay per output), measured jobs and memory checks, FAQ, and the monthly email.
- [How we stay independent](https://nodegrove.io/independence): how the site makes money, the referral programs it uses, the rules that keep verdicts honest, and how every number is measured.

## Free tools and reference

- [Can I run it?](https://nodegrove.io/tools/can-i-run-it): a GPU and a model in; fits or not, how fast, and what to change when it does not.
- [LLM VRAM calculator](https://nodegrove.io/tools/llm-vram-calculator): weights + KV cache + overhead for any open model, with the formula shown.
- [Tokens-per-second estimator](https://nodegrove.io/tools/llm-speed-estimator): a generation speed ceiling from memory bandwidth and weight size.
- [Build vs rent calculator](https://nodegrove.io/tools/build-vs-rent-calculator): the break-even month for buying a GPU against renting one.
- [ComfyUI VRAM checker](https://nodegrove.io/tools/comfyui-vram-checker): an open image or video model and an NVIDIA card in; which file loads fully, stage by stage, from exact file sizes.
- [ComfyUI workflow checker](https://nodegrove.io/tools/comfyui-workflow-checker): a ComfyUI workflow (.json, or a PNG ComfyUI saved) in; every model file it loads with its exact size, folder and download link, the VRAM at its largest stage on an NVIDIA card, the smaller files to swap in so it fits, the custom nodes it needs, and whether it runs on Comfy Cloud (its custom nodes checked against the last public copy of Comfy Cloud's node catalog, and its memory against Comfy Cloud's 96 GB card). Runs in the browser; the workflow is never uploaded.
- [Rent a GPU: cloud GPU prices](https://nodegrove.io/guides/cloud-gpu): on-demand prices per GPU-hour from RunPod, Lium, SaladCloud, Lambda, Hyperstack, Nebius, Jarvislabs, DigitalOcean, CoreWeave, Modal, AWS, Google Cloud and Azure, for every card from the RTX 4090 to the B200; a picker for the cheapest card that runs a given model, the cost per million tokens, and what stopped machines, storage and data transfer cost. Vast.ai is described without prices, since its terms do not allow them in a comparison.
- [ComfyUI in the cloud](https://nodegrove.io/guides/comfyui-cloud): every way to run ComfyUI in the cloud compared by what you pay: per-generation services (Comfy Cloud, Floyo, ComfyICU), rented ComfyUI machines (RunComfy, ThinkDiffusion, MimicPC) and raw GPUs (RunPod, and Vast.ai without its prices), with a session-cost calculator and the card each workflow needs.
- [ComfyUI API](https://nodegrove.io/guides/comfyui-api): how ComfyUI's own /prompt API works, and every way to run a workflow as an endpoint compared: Comfy API and the Comfy Cloud API, RunComfy, ViewComfy, fal, Replicate, RunPod Serverless and Modal, on price per GPU hour, what one job costs with idle time, custom nodes, cold starts, MCP and spend limits, regions and training terms.
- [Ollama Cloud](https://nodegrove.io/guides/ollama-cloud): Ollama Cloud's per-token pricing, what Pro and Max credits buy on each cloud model, which cloud models fit a home machine, the break-even speed for running them on your own card, and Ollama's privacy terms.
- [Image and video models](https://nodegrove.io/create): VRAM requirements in ComfyUI for Wan 2.2, FLUX.2, LTX-2.5, MiniMax H3, Qwen-Image and more, each with its own page.
- [Models](https://nodegrove.io/models): VRAM requirements of open models, each with its own page.
- [GPUs](https://nodegrove.io/gpus): which open models each GPU can run, each with its own page.
- [OpenClaw VRAM requirements](https://nodegrove.io/guides/openclaw-vram-requirements): the context OpenClaw and Hermes Agent need, its memory cost, and the best tool-calling local model per GPU.
- [Personal AI computers compared](https://nodegrove.io/guides/personal-ai-computer): Ghost Core, DGX Spark, RTX Spark laptops, Ryzen AI Max+ 395 mini PCs, Mac Studio M5 Ultra, an RTX 5090 PC and an RTX PRO 6000 workstation, compared on the memory the GPU can use, bandwidth and the models each one runs, including the best tool-calling model at 64k agent context. Computed from makers' specs, not a hands-on review.
- [DGX Spark vs Strix Halo vs Mac Studio](https://nodegrove.io/guides/dgx-spark-vs-strix-halo-vs-mac-studio): the unified-memory machines compared for local LLMs, from makers' specs.
- [Guides](https://nodegrove.io/guides): private AI servers, self-hosted LLMs, what fits on 8 to 96 GB.
- [MCP server](https://nodegrove.io/mcp): the checker inside Claude, Cursor, VS Code and any MCP client, free and read-only. Remote endpoint https://mcp.nodegrove.io/mcp (Streamable HTTP, no auth) or local `npx -y @nodegrove/vram-mcp`; tools can_i_run, what_fits, estimate_vram, estimate_from_hf_repo (any Hugging Face repo), list_models, list_gpus. Source: https://github.com/nodegrove/vram-mcp

## Japanese (日本語)

- [日本語のトップ](https://nodegrove.io/ja): the home page in Japanese. The free tools, the model pages, the image and video model pages, the OpenClaw guide, the independence page and three cloud guides also exist in Japanese under /ja with the same paths (for example https://nodegrove.io/ja/models/qwen3.8-27b). Every figure is computed by the same code as the English page.
- [ComfyUIをクラウドで使う](https://nodegrove.io/ja/guides/comfyui-cloud): the ComfyUI cloud comparison in Japanese, adding GPUSOROBAN priced in yen with tax included.
- [ComfyUI APIの使い方とAPI化サービス比較](https://nodegrove.io/ja/guides/comfyui-api): the ComfyUI API comparison in Japanese.
- [Ollama Cloudの料金・モデル・制限・プライバシー](https://nodegrove.io/ja/guides/ollama-cloud): the Ollama Cloud guide in Japanese, adding its off-peak hours in Japan time.

## Open data

- [LLM VRAM dataset](https://nodegrove.io/data): memory for open models at six quantisations and every common context length, and how each fits the GPUs above. CSV and JSON under CC BY 4.0, readable by URL (https://nodegrove.io/data/llm-vram.json). Every version is a release at https://github.com/nodegrove/llm-vram-dataset and is archived on Zenodo; cite it by DOI 10.5281/zenodo.22966137 (all versions, resolves to the newest).
- [Cloud GPU prices (JSON)](https://nodegrove.io/data/cloud-gpu-prices.json): the cloud GPU guide's on-demand prices per GPU per hour, before tax, with each provider's billing, stopped-machine, storage, data-out and sign-up terms, the date each provider was read, how each price is taken (list price or cheapest listing) and every card's memory and bandwidth. Open CORS, CC BY 4.0 with credit to Nodegrove. Vast.ai's prices are left out because its terms do not allow them in a comparison; read them at https://vast.ai/pricing.

## Facts

- Measured: on 20 October 2026, in Finland, Z-Image Turbo (4 images), Wan 2.2 5B (a 5-second video) and a 70B LLM batch were run on an RTX PRO 6000 (96 GB) and an RTX 6000 Ada (48 GB), with GPU memory sampled every 250 ms. For LLMs the site's computed memory landed within 1% of the measured peaks.
- Privacy: the checkers read workflows and files in the browser; Nodegrove stores no workflows, prompts or outputs.
- Contact: info@nodegrove.io