benchloop-cli 0.2.3

Local-first CLI + web dashboard for benchmarking LLMs across quality, speed, reliability, and a real multi-turn agent loop. Hardware-aware, deterministic, reproducible.

BenchLoop

Benchmark local LLMs by what actually matters.

BenchLoop is a local-first CLI + web app for benchmarking LLMs running on your own hardware. It scores models across seven repeatable suites — quality, speed, reliability, agentic tool use, coding, instruction following — and gives you receipts: per-task outputs, latency, token counts, machine info, scores.

No accounts, no telemetry, no API keys. Your model, your machine, your numbers.

$ benchloop run --model qwen3:8b --suites speed,toolcall,agent
... 8 tasks, 4 tools, 6 turns avg, 74.6 tok/s ...

Overall  73.4  ████████░░
Quality  73.6  ████████░░
Speed    78.9  █████████░
Agent    96.9  █████████▌

Published runs live at https://bench-loop.com/leaderboard. Every completed local benchmark auto-publishes there.

Why

Hosted LLM leaderboards answer "which model wins on a server farm someone else paid for?" BenchLoop answers "which model + harness + hardware combination actually works for me right now?" — the question you have when picking a local stack.

It is repeatable on purpose: every run persists to disk, the task set is frozen, the scorer is deterministic. If you say "qwen3:8b scored 89 on my 4090", anyone can install BenchLoop and verify it.

Install

pipx (recommended)

pipx install benchloop-cli
benchloop --version

The PyPI distribution is named benchloop-cli (the bare benchloop name was taken by an unrelated dataset library). The installed commands are still benchloop and bench-loop.

pip

pip install benchloop-cli

From source

git clone https://github.com/outsourc-e/bench-loop
cd bench-loop
pip install -e .

Run your first benchmark

Make sure you have a local LLM endpoint running. Anything OpenAI-compatible or Ollama-flavored works:

• Ollama at http://localhost:11434 (default)
• LM Studio at http://localhost:1234 (--provider openai_compat)
• MLX / Osaurus at http://localhost:8000 (--provider openai_compat)
• vLLM, Jan, llama-server, etc.

Then:

benchloop run \
  --model qwen3:8b \
  --endpoint http://localhost:11434 \
  --provider ollama

This runs every default suite, scores them, prints a console report, and persists the full run to ~/.bench-loop/runs/.

Run a subset

benchloop run --model qwen3:8b --suites speed,agent

Different prompting harness

Same model, four ways to talk to it:

benchloop run --model qwen3:8b --harness raw      # native tool calling
benchloop run --model qwen3:8b --harness hermes   # <tool_call>{...}</tool_call>
benchloop run --model qwen3:8b --harness qwen     # <function_call>{...}</function_call>
benchloop run --model qwen3:8b --harness pi       # <think>...</think> + Hermes tags

Stamp custom hardware (e.g. when benchmarking through a tunnel)

benchloop run \
  --model qwen3:8b \
  --endpoint http://localhost:11435 \
  --hardware "NVIDIA RTX 4090 24GB" \
  --gpu "NVIDIA RTX 4090" \
  --gpu-memory-gb 24

Launch the local dashboard

v0.2.0+ ships the full FastAPI + React dashboard inside the wheel. After pipx install benchloop-cli:

benchloop dashboard
# → open http://127.0.0.1:8877

This serves the Models, Benchmark, Leaderboard, Compare, and Chat tabs on a single port, with auto-discovered local providers (Ollama, LM Studio, MLX/Osaurus, vLLM, Jan).

For hot-reload development against a clone of bench-loop-web:

benchloop dashboard --dev

Suites

Suite	What it scores
speed	Latency, throughput, TTFT, generation tok/s across short/medium/long contexts
toolcall	Structured tool-call correctness across realistic tasks (weather, stocks, email, search)
coding	Executable Python tasks verified in a sandboxed subprocess (10s timeout)
dataextract	JSON / structured extraction from messy natural language
instructfollow	Constraint following, formatting, exactness
reasonmath	Small reasoning + math tasks with deterministic checks
agent	Multi-turn agentic tool use. BenchLoop drives a real loop: model emits a tool call, BenchLoop executes it locally, feeds the result back, model iterates until done. Scores correctness, efficiency, no-hallucination, required-tool coverage.

Scoring

Overall = 0.55 · quality + 0.20 · speed + 0.25 · reliability

• Quality = mean of non-speed suite scores (size-fair).
• Speed = 12.54 · log2(tok/s) + 0.9, clamped to 0–100.
• Reliability = pass rate across all tasks.
• Agent = correct_final + efficient + no_hallucinated_tools + all_required_called, 25 pts each, averaged across tasks.

Local web app

A FastAPI backend + React frontend bundle ships alongside the CLI for visualizing runs:

benchloop dashboard   # starts the local web app on :5180

Tabs: Models, Benchmark, Leaderboard, Compare runs, Chat, agent trace viewer.

Publish a run

Every completed benchmark auto-publishes to https://bench-loop.com/leaderboard via https://api.bench-loop.com/submit. Runs are deduped by (machine_id, run_id) so the same run from the same machine won't be double-counted.

Opt out:

export BENCHLOOP_NO_SUBMIT=1

You can still manually export a snapshot for sharing / archiving:

benchloop export --output my-runs.json

Architecture

bench-loop/                    ← this repo, the CLI + suites + scorers
  bench_loop/
    cli.py                     ← `benchloop` entrypoint
    suites/                    ← speed, toolcall, coding, agent, ...
    harness.py                 ← raw / hermes / qwen / pi adapters
    providers/                 ← ollama, openai_compat
    runner/orchestrator.py     ← drives suites + harnesses
    tasks/                     ← frozen task YAML fixtures
bench-loop-web/                ← the web app (separate repo)
  api/                         ← FastAPI wrapper around bench_loop
  ui/                          ← local dashboard
  site/                        ← public bench-loop.com static site

Status

BenchLoop is v0.1 beta. The benchmark surface, scoring, web app, agent loop, and four harnesses all work end-to-end. Stuff still on the roadmap:

• Streaming TTFT for OpenAI-compatible providers (currently 0 on those backends — ollama TTFT is fine)
• Bigger task fixtures (each suite is intentionally small and frozen for v1)
• Hosted submission flow for community runs
• More provider adapters (TGI, Bedrock, etc. if there's demand)

License

MIT. See LICENSE.