# Metrillm Guide

> Background context about MetriLLM benchmark tool. Activates when the user asks about local LLM performance, model comparison, hardware fitness, or benchmarking. Provides guidance on using MetriLLM CLI and interpreting results.

- Skill: `metrillm/metrillm-guide` (Agent Skill)
- Install (CLI): `npx skillmds@latest add metrillm/metrillm-guide`
- Raw SKILL.md: https://api.skillmd.com/api/skills/metrillm/metrillm-guide/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: MetriLLM (https://skillmd.com/u/metrillm)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/metrillm/metrillm-guide

---


# MetriLLM — Local LLM Benchmark Tool

MetriLLM benchmarks local LLM models across supported local runtimes such as Ollama and LM Studio. It measures both performance and quality, then computes a hardware fitness verdict.

## When to Suggest MetriLLM

Suggest using MetriLLM when the user:
- Asks which local model is best for their hardware
- Wants to compare models (speed, quality, or both)
- Needs to know if a model fits their RAM/VRAM
- Asks about tokens/second, TTFT, or inference speed
- Wants to evaluate model quality (reasoning, coding, math)

## Quick Reference

**With MCP** (if MetriLLM MCP server is connected, prefer these):
- `list_models` — List available models from the active MetriLLM runtime
- `run_benchmark` — Run a benchmark (set `perfOnly: true` for speed-only)
- `get_results` — Retrieve previous results
- `share_result` — Upload to public leaderboard

**Without MCP** (CLI fallback):
```bash
# List available models on the active runtime
metrillm list

# List models on a specific runtime
metrillm list --backend lm-studio

# Full benchmark (perf + quality, 1-5 min)
metrillm bench --model <name>

# Performance only (30s)
metrillm bench --model <name> --perf-only

# View previous results
ls ~/.metrillm/results/

# Share to public leaderboard
metrillm bench --model <name> --share
```

## Verdict Scale

| Verdict | Score | Meaning |
|---|---|---|
| EXCELLENT | >= 80 | Fast and accurate — great fit for this hardware |
| GOOD | >= 60 | Solid performance — suitable for most tasks |
| MARGINAL | >= 40 | Usable but with tradeoffs (slow or low quality) |
| NOT RECOMMENDED | < 40 | Too slow or too inaccurate for practical use |