# Nodegrove VRAM: can I run it?

Works out whether an open-weight model fits in your card's memory, how much context survives, and what to run instead. Handy before you start a local model.

- MCP server · `type=mcp` · Claude Code, Codex, Cursor, GitHub Copilot, Windsurf (Devin Desktop), Cline, Zed, Gemini CLI
- Author: nodegrove
- License: MIT · policy=link-only · description only; the file stays with the author
- Source: https://github.com/nodegrove/vram-mcp
- Updated: 2026-10-19
- HTML: https://affpapa.org/en/nevibecode/mcp/nodegrove-vram-can-i-run-it

## What it does

Ask it something like whether an RTX 4090 handles Llama 3.3 70B at 32k, and it answers in fields: fits, tight or no, the split between weights, KV cache and overhead, a speed ceiling, and the longest context that still works. On a no it lists the changes that would make the model fit. Models go in as people type them or as any Hugging Face repo id; cards by name or by their memory and bandwidth.

Six tools sit behind it: `can_i_run`, `what_fits`, `estimate_vram`, `estimate_from_hf_repo`, `list_models` and `list_gpus`. A foreign repo is read from its `config.json`. Anything the maths cannot cover gets named in the answer rather than guessed at, and every tool only reads. Memory is counted per architecture: sliding-window layers cap their cache at the window, hybrid models grow one on the few full-attention layers, so a plain transformer formula would overshoot badly at long context.

The remote server lives at `https://mcp.nodegrove.io/mcp` and asks for no account or key. For a local one use `npx -y @nodegrove/vram-mcp`, which needs Node.js 20 or newer. The speed figure is a single-stream ceiling, not a benchmark; real runtimes come in lower, and the higher the number the wider the gap.

## Requirements
- Needs: Node.js (commands run via npx)
- Remote server: `https://mcp.nodegrove.io/mcp`

## Install

### Claude Code (native platform)

1. **Platform CLI: remote server**
```bash
claude mcp add --transport http vram-mcp https://mcp.nodegrove.io/mcp
```

2. **Platform CLI**
```bash
claude mcp add nodegrove-vram-can-i-run-it -- npx -y @nodegrove/vram-mcp
```

### Codex

1. **Platform CLI: remote server**
```bash
codex mcp add nodegrove-vram-can-i-run-it --url https://mcp.nodegrove.io/mcp
```

2. **Platform CLI**
```bash
codex mcp add nodegrove-vram-can-i-run-it -- npx -y @nodegrove/vram-mcp
```

### Cursor

**Config file**
File `.cursor/mcp.json`:
```json
{
    "mcpServers": {
        "nodegrove-vram-can-i-run-it": {
            "url": "https://mcp.nodegrove.io/mcp"
        }
    }
}
```
File `.cursor/mcp.json`:
```json
{
    "mcpServers": {
        "nodegrove-vram-can-i-run-it": {
            "command": "npx",
            "args": [
                "-y",
                "@nodegrove/vram-mcp"
            ]
        }
    }
}
```

### GitHub Copilot

**Config file**
File `.vscode/mcp.json`:
```json
{
    "servers": {
        "nodegrove-vram-can-i-run-it": {
            "type": "http",
            "url": "https://mcp.nodegrove.io/mcp"
        }
    }
}
```
File `.vscode/mcp.json`:
```json
{
    "servers": {
        "nodegrove-vram-can-i-run-it": {
            "type": "stdio",
            "command": "npx",
            "args": [
                "-y",
                "@nodegrove/vram-mcp"
            ]
        }
    }
}
```

### Gemini CLI

1. **Platform CLI: remote server**
```bash
gemini mcp add -t http nodegrove-vram-can-i-run-it https://mcp.nodegrove.io/mcp
```

2. **Platform CLI**
```bash
gemini mcp add nodegrove-vram-can-i-run-it npx -y @nodegrove/vram-mcp
```

## When to use it
- Check whether a model fits your card before downloading the weights
- Pick a context length that stays inside VRAM
- Find a replacement for a model that does not fit
- Compare memory cost across quantisations
- Estimate a card that is not in the ready-made tables

## Example prompts
- Will Qwen3 32B fit on 24 GB at 64k context?
- What should I run instead of Llama 3.3 70B on a 16 GB card?
- How much memory does `Qwen/Qwen3-Next-80B-A3B-Instruct` take at Q4_K_M?

## Similar
- [hf-mem](https://affpapa.org/en/nevibecode/skills/hf-mem.md) — skill · 858 installs
- [FitLLM](https://affpapa.org/en/nevibecode/mcp/fitllm.md) — MCP server · ★ 8
- [discovery](https://affpapa.org/en/nevibecode/skills/discovery.md) — skill · 2600 installs
- [run-experiment](https://affpapa.org/en/nevibecode/skills/run-experiment.md) — skill · 523 installs
- [huggingface-vision-trainer](https://affpapa.org/en/nevibecode/skills/huggingface-vision-trainer.md) — skill · 1900 installs
- [tao-run-on-kubernetes](https://affpapa.org/en/nevibecode/skills/tao-run-on-kubernetes.md) — skill · 1600 installs

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md