DEDUPED Hot AI repos, turned into use cases Get the Monday 5

How to find why a local model slowed down with Mlxtop

Read memory, paging, GPU and runtime speed side by side to see which one changed.

Short answer: Install Mlxtop with brew install maximpri/tap/mlxtop, or use the install script from the README on Linux and macOS; then start your runtime and send it a prompt so a model is loaded; run mlxtop to open the dashboard. All steps.

Who it is forA person running a local model on their own Mac or PC who notices replies have become slow and wants to know whether the machine or the model is at fault.

The problemA slow local model can come from memory pressure, swapping, a full GPU or a throttled card, and the runtime's own logs rarely show this.

The outcomeA single verdict line on the Overview screen names the likely cause, backed by memory, paging, GPU and tokens per second readings.

What you need first

  • macOS 11 or later on Apple Silicon, Linux (x86_64 or ARM64), or Windows (x86_64)
  • A local LLM runtime such as oMLX, Ollama, llama.cpp or vLLM installed and able to serve a model
  • A terminal, or Rust 1.88 or newer if you build from source

Steps

  1. Install Mlxtop with brew install maximpri/tap/mlxtop, or use the install script from the README on Linux and macOS.
  2. Start your runtime and send it a prompt so a model is loaded.
  3. Run mlxtop to open the dashboard.
  4. Watch view 1 · Overview while the model answers and read the verdict line.
  5. Press d for diagnostics and setup help for the current runtime if a reading is missing.

Prompt for your agent

Using the mlxtop repo at maximpri/mlxtop, run `mlxtop --once` on <target machine> while the local model served at <endpoint> is responding slowly. Report the verdict line and whether it names memory pressure, paging, a model partly on the CPU, a full card or heat throttling. Verify the report includes tokens per second and memory figures before you state a cause.

Caveat. Mlxtop shows what changed when the model slowed down but cannot prove the cause, and GPU activity covers the whole system rather than one model.

Use this in your AI agent

Ask your agent

Using the Deduped MCP, get the setup steps for maximpri/mlxtop and "Find why a local model slowed down"

Add Deduped to Claude Code once

claude mcp add --transport http deduped https://dedupedai.com/mcp

Cursor, ChatGPT, others

Share: X LinkedIn Hacker News Email Copy link

Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer