Show HN: mlx-chronos - benchmark MLX inference engines on Apple Silicon

Benchmark suite and community leaderboard for local LLM inference on Apple Silicon. Run a reproducible benchmark, save a sealed JSON result, and compare engines across Macs.

OverviewSupported EnginesQuick StartCLI ReferenceConfigurationBenchmark ProtocolLeaderboard RulesSubmit ResultsRoadmap

mlx-chronos

is a standardized benchmark tool for local LLM inference engines on Apple Silicon. It detects your Mac, runs a fixed benchmark protocol against an OpenAI-compatible engine endpoint, and writes structured result files for local analysis or public leaderboard submission.

The public leaderboard is available at igurss.github.io/mlx-chronos.

Metric	Meaning	Public comparison use
TTFT cold	Time from request start to first non-empty streamed token with cache-avoiding prompts	Yes
TTFT cached	Time to first token after a cache-priming call with the same prompt	Yes
Request throughput	Completion tokens divided by full client-observed request time	Yes, when engine token usage is reliable
Sustained throughput	Optional long throughput run for heat buildup and late-run degradation	Yes, under the sustained profile
System RAM peak	Peak total Mac RAM in use during the benchmark	Yes
Engine RSS	Post-warmup RSS of the engine server process when identifiable	Diagnostic only
Thermal state	Start, end, worst state, samples, and affected benchmark phases when available	Context metadata
Tool calling	Planned future success-rate benchmark	Not yet available

0.3.1

simplifies public model identity metadata to model name, quantization, model format, and the required model reference URL, while keeping the guided workflows, timing metadata, and stricter leaderboard integrity checks introduced in 0.3.0

.

Engine	Project	Notes
Ollama

jundot/omlxraullenchai/Rapid-MLXwaybarrios/vllm-mlxml-explore/mlx-lm

NoteThe engine server must already be running beforemlx-chronos run

,mlx-chronos models

, ormlx-chronos validate

can query it. See[CONTRIBUTING.md]for engine setup details.

pip install mlx-chronos

Optional thermal-state support through macOS Foundation/PyObjC:

pip install "mlx-chronos[thermal]"
mlx-chronos --version
mlx-chronos upgrade

When run in an interactive terminal, mlx-chronos

performs a best-effort background PyPI version check. If a newer release is available, it prints a short notice recommending:

mlx-chronos upgrade

Set MLX_CHRONOS_DISABLE_UPDATE_CHECK=1

to disable the automatic check.

mlx-chronos engines
mlx-chronos models --engine omlx
mlx-chronos validate --engine omlx --model "Qwen3.5-4B-OptiQ-4bit"
mlx-chronos wizard

The wizard provides a terminal menu for common actions and a guided benchmark builder with engine, model, profile, token bounds, output format, cooldown, preflight, notes, and other run options. When the selected engine server is running, the wizard loads /models

and lets you select a model from the exposed IDs, with manual entry as a fallback. Before launching a benchmark, it shows the equivalent mlx-chronos run ...

command so the same configuration can be reused in scripts. You can return to the main menu from benchmark setup without starting a run.

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit"

Results are written to results/local/

by default.

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --format all

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --output-dir ~/Desktop/benchmarks

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --max-tokens 100 --min-tokens 80

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --profile sustained

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --cooldown-seconds 300

mlx-chronos run --engine omlx --model "Qwen3.5-4B-OptiQ-4bit" --preflight

mlx-chronos run --engine omlx \
  --model "Qwen3.5-4B-OptiQ-4bit" \
  --model-url "https://huggingface.co/mlx-community/Qwen3.5-4B-OptiQ-4bit"

Command	Purpose
mlx-chronos --version
Print the installed package version
mlx-chronos wizard
Open an interactive menu for common commands and guided benchmark setup
mlx-chronos upgrade
Check PyPI and upgrade the current Python environment if a newer release exists
mlx-chronos engines
List supported engines and local installed/running status
mlx-chronos models --engine <name>
List model IDs exposed by a running engine server
mlx-chronos validate --engine <name> --model <model>
Validate hardware, engine, server, and optional model access
mlx-chronos run --engine <name> --model <model>
Run a benchmark and save local result files
mlx-chronos submit --file <result.json> --dry-run
Validate whether a result is publishable
mlx-chronos submit --file <result.json>
Send a validated result to the maintainer inbox

Setting	Example	What it changes
MLX_CHRONOS_<ENGINE>_PORT
MLX_CHRONOS_OMLX_PORT=8002
Overrides an engine server port
MLX_CHRONOS_CACHED_TTFT_RATIO
MLX_CHRONOS_CACHED_TTFT_RATIO=0.8
Sets the cached-TTFT warning threshold
MLX_CHRONOS_DISABLE_UPDATE_CHECK
MLX_CHRONOS_DISABLE_UPDATE_CHECK=1
Disables automatic background update checks
MLX_CHRONOS_SUBMIT_ENDPOINT
https://example.test/form
Overrides the maintainer inbox endpoint

Default engine ports:

Engine	Default port
oMLX	8000
Rapid-MLX	8001
vllm-mlx	8000
mlx-lm	8080
Ollama	11434

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.

mlx-chronos run

executes a fixed protocol against the running engine. The JSON result records exact prompt text, token bounds, benchmark profile, timing metadata, hardware metadata, and an integrity seal.

Phase	What happens
Hardware detection	Captures chip, machine model, memory, macOS, Python, architecture, battery state, Low Power Mode, and thermal context when available
Warmup	Uses a separate prompt so same-run prefix/KV cache hits do not remove throughput prefill work
Cold TTFT	Uses unique prompts inside the run to avoid same-run cache hits
Cached TTFT	Primes one fixed prompt, then measures consecutive cached trials
Throughput	Uses fixed protocol prompts and deterministic generation parameters
RAM and thermal tracking	Samples system RAM, diagnostic engine RSS, phase timings, and thermal state where available
Result sealing	Adds a tamper-evident integrity seal for public-submission validation

• Requests use deterministic generation parameters: temperature=0.0

andtop_p=1.0

. - Throughput is end-to-end request throughput, not pure decode speed. It includes request overhead, prefill, and decode.

• Timed TTFT and throughput requests are never retried. A transient request failure invalidates the run instead of becoming part of a published timing.
• Cached TTFT is recorded only after cache priming completes successfully.
• Decode throughput records first-content-to-stream-end elapsed time so the value can be reconstructed from raw completion-token counts.
• Throughput prompts intentionally vary to reduce cache artifacts, so run standard deviation includes workload variation plus system and engine noise.
• If an engine cannot provide reliable usage.completion_tokens

, the run falls back to a local estimate and is marked as not leaderboard-comparable. - p95 is reported only when at least 20 trials are available.

• The default baseline run uses 5 trials. The maximum prompt pool supports 30 unique cold and throughput prompts.

--profile sustained

runs one 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.

If the sustained run observes a thermal-state change or non-nominal thermal state, result metadata includes a sustained throttling warning. The warning compares early and late progress-window averages, not a single first/last sample.

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

.

Use --cooldown-seconds

to enforce a before starting another run. The default recent-run warning threshold is 300 seconds.

For a fuller explanation, see docs/methodology.md.

Local runs are intentionally flexible. You can change trial count, profile, output token bounds, cooldown, connection mode, notes, and other parameters for your own diagnostics.

Public leaderboard submissions are stricter so rows remain comparable.

| Profile | Trials | max_tokens | Minimum generated output | min_tokens | |---|---|---|---|---| | Baseline | 5 | 100 | 80 tokens | Not allowed | | Sustained | 1 | 1000 | 800 tokens | Not allowed |

• Throughput must use the engine response's usage.completion_tokens

. - The result must include model.reference_url

, a link to the model used. - The inference engine version must be known; engine.version=unknown

is not accepted for public comparison. - Hardware must report an Apple M-series chip, arm64

, and a valid macOS version; timestamps may not be more than 10 minutes in the future. - All warmup calls must complete successfully ( warmup_failures=0

). - System RAM, engine RSS, and continuous Foundation thermal monitoring must complete without sampling errors.

• macOS Low Power Mode must be disabled.
• Decode throughput must include reconstructible raw decode elapsed time.
• The JSON must pass mlx-chronos submit --dry-run

. - The result must include a valid integrity seal.

• The archive rejects duplicate integrity digests and duplicate run identities.
• Custom token bounds, fallback token estimates, custom public-profile trial counts, short-output runs, and Low Power Mode runs are valid local records but are not accepted into the public leaderboard.

Result JSON also contains internal benchmark-protocol labels 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. Model reference URLs point to the model page used for the run. Model pages can change over time when maintainers update files or tags. Leaderboard comparisons keep model name, quantization, format, and model reference URL separate so distinct variants are not grouped together.

Run

mlx-chronos run

on your Mac. - Find the generated JSON in

results/local/

. - Validate it locally:

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

Copy the checked JSON into

results/submitted/

with a clear filename. - Open a pull request with only that JSON file changed.

GitHub Actions labels the PR as

result-submission

, validates schema and integrity, and the maintainer reviews it before merge.

WarningDo not edit submitted JSON by hand after the run. Public submissions include anintegrity

seal over the canonical result payload; changing any benchmark field invalidates that seal.

If opening a PR is inconvenient, send a validated result directly:

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

Maintainers can override the inbox endpoint with --endpoint

or MLX_CHRONOS_SUBMIT_ENDPOINT

.

See CONTRIBUTING.md for detailed contributor instructions.

• 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

, anddocumentation

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

• CLI update notifications and mlx-chronos upgrade

• 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

Apache 2.0. See LICENSE.

source & further reading

github.com — original article