mlx-chronos 0.2.0

Benchmark suite and community leaderboard for local LLM inference on Apple Silicon

mlx-Chronos ⏱️

Benchmark suite and community leaderboard for local LLM inference on Apple Silicon.
Run it. Share your results. Compare across hardware.

---

What is mlx-Chronos?

mlx-Chronos is a standardized benchmarking tool for local LLM inference engines on Apple Silicon. It automatically detects your hardware, runs a consistent set of tests across installed engines, and produces a structured JSON result you can contribute to the community leaderboard.

Supported engines:

• Ollama (MLX backend)
• oMLX
• Rapid-MLX
• vllm-mlx
• mlx-lm (Apple MLX)

Metrics measured:

• TTFT — Time to First Token (cold and cached, with statistics)
• tok/s — Client-observed request throughput (mean, stddev, min, max across trials)
• Sustained tok/s — Optional long throughput profile for heat buildup and late-run degradation checks
• Output tokens — Completion token counts for throughput trials
• System RAM peak — Peak total Mac RAM in use during the benchmark, used as the public memory comparison metric
• Post-warmup engine RSS — Legacy diagnostic peak RSS of the engine server process when available, not a comparison metric
• Tool calling — Success rate (future work)

---

How It Works

When you run mlx-Chronos, it executes a fixed benchmark protocol against the running engine:

Cold TTFT — sends a prompt to the model and measures the time from request to first non-empty streamed token, including whitespace-only text tokens. Each trial uses a unique prompt inside the run to avoid same-run cache hits.

Cached TTFT — sends the same fixed prompt on every cached trial. A priming call loads it into cache first, then cached trials run consecutively. This measures cache performance without interleaving unrelated prompts between cached measurements. If cached TTFT is close to cold TTFT, new runs record a warning that cache reuse may not have happened.

Request throughput (tok/s) — measures completion tokens divided by the full client-observed request time using fixed protocol prompts, one per trial. Warmup uses a separate prompt so same-run prefix/KV cache hits do not silently remove throughput prefill work. Requests use deterministic generation parameters (temperature=0.0, top_p=1.0). This includes request overhead, prefill, and decode, so it is an end-to-end throughput metric rather than pure decode speed. Throughput prompts intentionally vary to reduce cache artifacts, so per-run stddev includes normal workload variation as well as system and engine noise. New runs also record client-observed decode_tokens_per_second from the streaming throughput trial when reliable completion-token usage is available. If an engine cannot provide usage in the streaming response, the run falls back to a local estimate and is marked as not leaderboard-comparable. Throughput uses a fixed requested max_tokens value by default, and optional output token bounds can be requested with --max-tokens / --min-tokens for local experiments.

System RAM peak — continuously samples total Mac RAM usage from before warmup through the recorded benchmark phases and reports the observed peak in GB and percent. This is the public leaderboard memory metric because it answers the practical question of how much memory pressure the run placed on the Mac.

Thermal monitor — records phase timings plus a lightweight thermal summary during the run when macOS thermal state is available through Foundation/PyObjC. New JSON results include start/end/worst thermal state, sample count, and phases where non-nominal thermal state was observed. Without PyObjC/Foundation, the continuous monitor is unavailable even if a one-shot pre-run thermal state can be read through powermetrics.

Sustained profile — --profile sustained runs a single long throughput trial with max_tokens=1000 by default and records progress samples every 100 generated output units. Intermediate samples are estimates when the stream only reports exact token usage at the end. These samples make late-run throughput drops easier to spot. When a sustained run also observes a thermal-state change or non-nominal thermal state, mlx-Chronos records a sustained throttling warning in result metadata. The warning compares early and late progress-window averages, not a single first/last sample.

Cooldown tracking — before each run, mlx-Chronos checks the latest prior JSON result in the same output directory. The elapsed time is saved as meta.elapsed_since_last_benchmark_seconds; --cooldown-seconds can enforce a pause before starting a new run. The default recent-run warning threshold is a 300-second heuristic.

Post-warmup engine RSS diagnostic — records the resident memory of the engine server process after warmup, through the recorded benchmark phases, when the process can be identified. This is retained for debugging only: it is not total model memory and not a public comparison metric, because macOS/Metal unified-memory accounting can vary across environments. The default RAM sampling interval is 50ms and can be changed with --ram-sample-interval.

All metrics are run over multiple trials and reported with mean, stddev, min, and max. p95 is added only when at least 20 trials are available. The default is 5 trials, with a maximum of 30 unique cold and throughput prompts. Results are saved as structured JSON in results/local/ by default. Maintainers publish reviewed JSON files into results/submitted/ after accepting them for the community leaderboard. Result JSON records benchmark protocol metadata, including exact prompt text and requested token bounds, so runs can be reproduced without digging through source code. New public submissions must also carry a tamper-evident integrity seal that GitHub Actions verifies before a row can enter the leaderboard.

---

Community Leaderboard

Compare submitted results by model, chip, and RAM:

→ igurss.github.io/mlx-chronos

The default view compares engines for a selected model and Mac configuration. The raw-data view keeps the submitted rows available with optional columns for useful comparison metadata such as profile, thermal state, exact machine model, engine version, and quantization.

---

Current Release

0.2.0 tightens public leaderboard comparability, adds vllm-mlx support, refreshes the leaderboard UI, and makes benchmark request semantics explicit in result metadata.

---

Quick Start

# Install
pip install mlx-chronos

# Confirm installed version
mlx-chronos --version

# Check available engines
mlx-chronos engines

# List models exposed by a running engine server
mlx-chronos models --engine omlx

# Validate setup before a run
mlx-chronos validate --engine omlx --model "Qwen3.5-4B-OptiQ-4bit"

# Run benchmark (JSON by default)
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit"

# Optional: request throughput output token bounds
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --max-tokens 100 --min-tokens 80

# Optional: sustained profile for a longer heat/throttling-sensitive run
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --profile sustained

# Optional: enforce a pause after a recent run in the same output directory
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --cooldown-seconds 300

# Optional: write both JSON and Markdown outputs
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --format all

# Optional: choose a custom output directory
mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --output-dir ~/Desktop/benchmarks

Note: the engine server must be running before you launch mlx-chronos. See CONTRIBUTING.md for setup instructions.

Optional thermal-state support through macOS Foundation can be installed with pip install "mlx-chronos[thermal]". Engine ports can be overridden with MLX_CHRONOS_<ENGINE>_PORT, for example MLX_CHRONOS_OMLX_PORT=8002, MLX_CHRONOS_VLLM_MLX_PORT=8003, or MLX_CHRONOS_MLX_LM_PORT=8002. oMLX and vllm-mlx both default to port 8000. To avoid mislabeling results, mlx-Chronos checks the oMLX listener process with lsof; if that process cannot be inspected, oMLX validation may fail even when /v1/models responds.

---

Local Runs vs Public Leaderboard

mlx-chronos run is intentionally flexible for local diagnostics. You can change trial count, profile, output token bounds, cooldown, connection mode, notes, and other parameters to investigate your own engines and hardware. Those local JSON files are still valid benchmark records.

The public leaderboard is stricter. Only results that pass mlx-chronos submit --dry-run are publishable, and GitHub Actions applies the same policy before adding a row to the leaderboard. That keeps public rows comparable while leaving the local tool useful for experiments.

Result JSON contains an internal benchmark-protocol label used by validators to detect incompatible result formats. Treat labels such as 1, 2, and 3 as implementation compatibility markers, not public protocol release versions. User-facing documentation describes the measurement behavior instead of asking contributors to reason about those labels.

Contributing Your Results

1. Run mlx-chronos run on your Mac
2. A JSON file is generated in results/local/ (use --format all for a Markdown summary too)
3. Check the result:

   mlx-chronos submit --file results/local/your-result.json --dry-run
   

4. Copy the checked JSON into results/submitted/ with a clear filename
5. Open a pull request with only that JSON file changed
6. GitHub Actions labels the PR as result-submission, validates the schema and integrity seal, and the maintainer reviews it before merge

Leaderboard submissions must report throughput using the engine response's usage.completion_tokens and keep one of the standard profiles: baseline with exactly 5 trials and max_tokens=100, or sustained with exactly 1 trial and max_tokens=1000. Neither profile may request min_tokens, and macOS Low Power Mode must be disabled. Each throughput trial must also generate at least 80% of the standard token limit: 80 tokens for baseline, 800 tokens for sustained. Custom local runs, fallback token estimates, custom token bounds, custom public-profile trial counts, short-output runs, or Low Power Mode runs can still be saved locally, but they are not accepted into the public leaderboard.

Do not edit submitted JSON by hand after the run. Public submissions include an integrity seal over the canonical result payload; changing any benchmark field invalidates that seal and the submission validator will reject the file.

If opening a PR is inconvenient, mlx-chronos submit --file ... still sends the validated JSON to the maintainer inbox as a fallback. Maintainers can override the inbox endpoint with --endpoint or the MLX_CHRONOS_SUBMIT_ENDPOINT environment variable.

See CONTRIBUTING.md for detailed instructions.

---

Benchmark Methodology

See docs/methodology.md for a full explanation of what is measured, how, and why.

---

Roadmap

Completed

• Core benchmark runner with repeated trials, warmup, cache priming, and phase-separated metrics
• Engine support for oMLX, Rapid-MLX, vllm-mlx, mlx-lm, and Ollama
• Hardware detection for chip, machine model, memory, macOS, Python, architecture, and thermal state
• Strict JSON schema validation with raw-trial consistency checks
• Continuous system RAM peak sampling, with post-warmup engine RSS kept as a diagnostic field
• Preflight validation for engine, server, and model access
• GitHub Actions validation for submitted results
• PR-based result submissions with automatic result-submission, code, and documentation labels
• GitHub Pages leaderboard with model/chip/RAM engine comparison and configurable raw-data columns
• JSON and Markdown result export
• mlx-chronos submit for sending validated JSON results to the maintainer inbox
• Warnings for battery mode, Low Power Mode, non-nominal thermal state, and unavailable thermal state
• Integration tests against mock OpenAI-compatible servers
• Larger fixed cold-prompt pool with optional p95 reporting for larger runs
• Request-throughput timing metadata and client-observed streaming decode throughput
• Phase timing metadata and lightweight continuous thermal monitoring
• Sustained benchmark profile, cooldown metadata, and strict local-vs-public leaderboard policy
• Public submission trust model with lightweight anti-spoofing checks
• External contributor workflow for code PRs and leaderboard result submissions

Future

• Evaluate a clearer TTFT naming model without breaking the v0.1 JSON contract
• Add tool-calling success-rate benchmarks
• Collect more results from M3, M4, and M5 systems

---

License

Apache 2.0 — see LICENSE