MetriLLM

Benchmark your local LLM models in one command. Speed, quality, hardware fitness — with a shareable score and public leaderboard.

Think Geekbench, but for local LLMs on your actual hardware.

npm install -g metrillm@latest
metrillm bench

What You Get

• Performance metrics: tokens/sec, time to first token, memory usage, load time
• Quality evaluation: reasoning, coding, math, instruction following, structured output, multilingual (14 prompts, 6 categories)
• Global score (0-100): 30% hardware fit + 70% quality
• Verdict: EXCELLENT / GOOD / MARGINAL / NOT RECOMMENDED
• One-click share: --share uploads your result and gives you a public URL + leaderboard rank

Real Benchmark Results

From the public leaderboard — all results below were submitted with metrillm bench --share.

Key takeaway: Small models (1-4B) fly on Apple Silicon. Larger models (14B+) with thinking chains can choke even on capable hardware. See full leaderboard →

Install

Requires Node 20+ and a local runtime: Ollama or LM Studio.

# Install globally
npm install -g metrillm@latest
metrillm bench

# Alternative package managers
pnpm add -g metrillm@latest
bun add -g metrillm@latest

# Homebrew
brew install MetriLLM/metrillm/metrillm

# Or run without installing
npx metrillm@latest bench

Usage

# Interactive mode — pick models from a menu
metrillm bench

# Benchmark a specific model
metrillm bench --model gemma3:4b

# Benchmark with LM Studio backend
metrillm bench --backend lm-studio --model qwen3-8b

# Benchmark all installed models
metrillm bench --all

# Share your result (upload + public URL + leaderboard rank)
metrillm bench --share

# CI/non-interactive mode
metrillm bench --ci-no-menu --share

# Force unload after each model (useful for memory isolation)
metrillm bench --all --unload-after-bench

# Export results locally
metrillm bench --export json
metrillm bench --export csv

Upload Configuration (CLI + MCP)

By default, production builds upload shared results to the official MetriLLM leaderboard (https://metrillm.dev).

• No CI secret injection is required for standard releases.
• Local/dev runs use the same default behavior.
• Self-hosted or staging deployments can override endpoints with:
  • METRILLM_SUPABASE_URL
  • METRILLM_SUPABASE_ANON_KEY
  • METRILLM_PUBLIC_RESULT_BASE_URL

If these variables are set to placeholder values (from templates), MetriLLM falls back to official defaults.

Windows Users

PowerShell's default execution policy blocks npm global scripts. If you see PSSecurityException or UnauthorizedAccess when running metrillm, run this once:

Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

Alternatively, use npx metrillm@latest which bypasses the issue entirely.

Runtime Backends

Shared runtime env:

• METRILLM_STREAM_STALL_TIMEOUT_MS (optional): stream watchdog for all backends, default 30000, 0 disables it

LM Studio benchmark runs now use the native REST inference endpoint (/api/v1/chat) for both streaming and non-streaming generation. The previous OpenAI-compatible inference path (/v1/chat/completions) has been retired from MetriLLM so tok/s and TTFT can rely on native LM Studio stats when available. If a LM Studio response omits native token stats, MetriLLM still computes a score and shows the throughput as estimated.

For very large models, tune timeout flags:

• --perf-warmup-timeout-ms (default 300000)
• --perf-prompt-timeout-ms (default 120000)
• --quality-timeout-ms (default 120000)
• --coding-timeout-ms (default 240000)
• --stream-stall-timeout-ms (default 30000, 0 disables stall timeout for any backend)

Benchmark Profile v1 (applied to all benchmark prompts):

• temperature=0
• top_p=1
• seed=42
• thinking follows your benchmark mode (--thinking / --no-thinking)
• Context window stays runtime default (context=runtime-default) and is recorded as such in metadata.

LM Studio non-thinking guard:

• When benchmark mode requests non-thinking (--no-thinking or default), MetriLLM now aborts if the model still emits reasoning traces (for result comparability).
• To disable it in LM Studio for affected models, put this at the top of the model chat template: {%- set enable_thinking = false %} then eject/reload the model.

How Scoring Works

Hardware Fit Score (0-100) — how well the model runs on your machine:

• Speed: 50% (tokens/sec relative to your hardware tier)
• TTFT: 20% (time to first token)
• Memory: 30% (RAM efficiency)

Quality Score (0-100) — how well the model answers:

• Reasoning: 20pts | Coding: 20pts | Instruction Following: 20pts
• Structured Output: 15pts | Math: 15pts | Multilingual: 10pts

Global Score = 30% Hardware Fit + 70% Quality

Hardware is auto-detected and scoring adapts to your tier (Entry/Balanced/High-End). A model hitting 10 tok/s on a 8GB machine scores differently than on a 64GB rig.

Full methodology →

Share Your Results

Every benchmark you share enriches the public leaderboard. No account needed — pick the method that fits your workflow:

All methods produce the same result:

• A public URL for your benchmark
• Your rank: "Top X% globally, Top Y% on [your CPU]"
• A share card for social media
• A challenge link to send to friends

Compare your results on the leaderboard →

MCP Server

Use MetriLLM from Claude Code, Cursor, Windsurf, or any MCP client — no CLI needed.

# Claude Code
claude mcp add metrillm -- npx metrillm-mcp@latest

# Claude Desktop / Cursor / Windsurf — add to MCP config:
# { "command": "npx", "args": ["metrillm-mcp@latest"] }

Full MCP documentation →

Skills

Slash commands that work inside AI coding assistants — no server needed, just a Markdown file.

Skills are included in the plugins below, or can be installed standalone:

# Claude Code
cp -r plugins/claude-code/skills/* ~/.claude/skills/

# Cursor
cp -r plugins/cursor/skills/* ~/.cursor/skills/

Plugins

Pre-built bundles (MCP + skills + agents) for deeper IDE integration.

Install:

# Claude Code
cp -r plugins/claude-code/.claude/* ~/.claude/

# Cursor
cp -r plugins/cursor/.cursor/* ~/.cursor/

See Claude Code plugin and Cursor plugin for details.

Integrations