Can I run it? in your AI assistant

Ask your assistant whether a model runs on your card and it answers with the arithmetic behind every figure on this site: the same formulas as the checker, over the same 40 models and 40 GPUs, and over any other model on Hugging Face, read from its config.json. It is free, read-only and needs no account or key.

Connect

Remote, with nothing to install. Add this URL to any client that speaks Streamable HTTP:

https://mcp.nodegrove.io/mcp

claude mcp add --transport http nodegrove-vram https://mcp.nodegrove.io/mcp

{
  "mcpServers": {
    "nodegrove-vram": {
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

{
  "servers": {
    "nodegrove-vram": {
      "type": "http",
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

devin mcp add -s user nodegrove-vram https://mcp.nodegrove.io/mcp

{
  "mcpServers": {
    "nodegrove-vram": {
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

{
  "mcpServers": {
    "nodegrove-vram": {
      "type": "streamableHttp",
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

codex mcp add nodegrove-vram --url https://mcp.nodegrove.io/mcp

gemini mcp add -s user --transport http nodegrove-vram https://mcp.nodegrove.io/mcp

{
  "context_servers": {
    "nodegrove-vram": {
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

Or run it on your own machine over stdio. It needs Node.js 20 or newer and nothing else:

{
  "mcpServers": {
    "nodegrove-vram": {
      "command": "npx",
      "args": [
        "-y",
        "@nodegrove/vram-mcp"
      ]
    }
  }
}

What you can ask

Six tools. You ask in your own words; the assistant picks the tool. Names are matched the way people type them, so "4090", "M4 Max" and "Llama-3.3-70B-Instruct" all work, and a card that is not listed can be described by its memory.

An answer

Asked whether an RTX 4090 runs Llama 3.3 70B, the server answers:

No: Llama 3.3 70B at Q4_K_M with 8,192 tokens of context needs 45.8 GB, and the RTX 4090 holds 22.8 GB after headroom. Short by 23 GB.

What would work instead:

• No context length helps: the weights alone are 43.1 GB before a single token of conversation.
• RTX 6000 Ada (48 GB) is within a whisker: 45.8 GB against 45.6 GB after headroom. With the KV cache at Q8 it needs 44.4 GB and fits.
• The smallest card here that runs it exactly as asked: A100, 80 GB usable.
• The biggest model your card does run at these settings, counting mixture-of-experts models at their dense equivalent: Gemma 4 31B, 20.9 GB at ~39 tokens/s.

Alongside the sentences come the numbers as fields (memory split, budget, speed ceiling, longest context), links to the model and card pages here, a link that opens the checker on the same settings, and the assumptions behind each figure. When a model does not fit, the answer also names the smallest common cloud GPU size that holds it, and where to run it covers the options.

How it works

The server is this site's arithmetic, published as the open-source package @nodegrove/llm-math. The pages, the calculators, the open dataset and the server all import it, so they cannot disagree, and a correction reaches all of them at once.

• Memory is weights plus KV cache plus overhead, with each model's cache counted the way it actually caches: sliding-window, hybrid and latent attention store far less than a standard transformer. The method has every constant.
• Models in the list were read from their config.json and checked against Hugging Face; the data version is 2026-11-03. Any other repo is read live by the same rules, and anything the reader cannot model is named in the answer rather than guessed.
• Speed is a ceiling from memory bandwidth, not a benchmark. The faster the figure, the further real runtimes fall below it, so figures of 100 tokens/s or more say "at most".

Limits

• Every figure is an estimate from stated formulas. Real usage moves with the runtime, batch size, flash attention and cache quantisation: read "fits" as "worth trying".
• For a mixture-of-experts model read from Hugging Face, the speed needs the active parameters from its model card; without them the answer gives memory only.
• The remote endpoint answers up to 120 requests a minute from one address. A Hugging Face lookup waits up to 8 seconds.

Privacy

• No account, no key, no cookies. The remote server runs on Cloudflare Workers. We keep no request logs and no record of what you ask: every request is answered by a fresh instance and forgotten. Cloudflare keeps standard edge logs for a short period, as for any website.
• The rate limit counts requests per address for one minute, inside Cloudflare, and keeps nothing after that.
• When you ask about a Hugging Face repo, the server fetches that repo's public config.json and metadata. Only the repo name is sent.
• The local version sends nothing anywhere, except to huggingface.co when you ask about a repo there.

The privacy notice has the same in full.

Open source

The server and the package are on GitHub at nodegrove/vram-mcp, the server is on npm as @nodegrove/vram-mcp, and it is listed in the official MCP Registry as io.nodegrove/vram-mcp. The code is MIT; the model and GPU data is CC BY 4.0, the same licence as the dataset. To use the math in your own code:

npm install @nodegrove/llm-math

Found an answer that disagrees with its source? Open an issue or write to info@nodegrove.io.

• Protocol: the Model Context Protocol, served to clients on the 2025 revisions and on 2026-07-28.
• Model architecture: each model's config.json on Hugging Face, linked from every answer and every model page.
• GPU memory and bandwidth: the manufacturers' specifications, linked from every GPU page.