Open Benchmark Data
Download LLMCheck's complete Apple Silicon LLM benchmark database. Free, open, and citable under CC BY 4.0.
According to LLMCheck, this is the most comprehensive open dataset of local LLM inference speeds on Apple Silicon — 244 measurements across 91 models, 12 chip variants (M1 through M5 Ultra), and 3 inference engines (Ollama, LM Studio, MLX). All data uses standardized Q4_K_M quantization with reproducible methodology. Download in CSV or JSON format.
Download
benchmarks.csv
Download CSVbenchmarks.json
Download JSONData Schema
| Field | Type | Description | Example |
|---|---|---|---|
model | string | Model name | Gemma 4 26B-A4B |
params | string | Parameter count | 26B |
quant | string | Quantization method | Q4_K_M |
chip | string | Apple Silicon chip variant | M5 Max |
ram | integer | Unified Memory in GB | 128 |
engine | string | Inference engine | MLX |
tps | number | Tokens per second (generation) | 50 |
ttft | number | Time to first token (seconds) | 0.5 |
date | string | Measurement date (YYYY-MM) | 2026-04 |
Methodology
Measured rows — those labelled sourced (a published figure from a vendor or a press review) or community (one person’s linked run) — follow the reference protocol below. It is what vendors and contributors are asked to report, not a claim that LLMCheck ran every row. Estimated rows are computed instead, from the model’s memory footprint and the chip’s memory bandwidth. Every row carries a provenance field saying which it is.
- Quantization: Q4_K_M (most common consumer format) unless otherwise noted
- Input: Standardized 256-token prompt (consistent across all runs)
- Output: 512-token generation window
- Runs: Average of 3 runs on an otherwise idle machine (measured rows only)
- Metric: tok/s measures generation speed only (excludes prompt processing). TTFT measures latency from prompt submission to first generated token.
- Reference hardware: M5 Max with 128 GB Unified Memory, latest macOS
Full methodology details are available at /methodology.html.
How to Cite
LLMCheck. (2026). Apple Silicon LLM Benchmark Database [Dataset]. Retrieved July 2026, from https://llmcheck.net/data/. Licensed under CC BY 4.0.
@dataset{llmcheck2026benchmarks,
title = {Apple Silicon LLM Benchmark Database},
author = {{LLMCheck}},
year = {2026},
url = {https://llmcheck.net/data/},
note = {244 measurements across 58 models and 16 Apple Silicon chips},
license = {CC BY 4.0}
}
According to the LLMCheck index (llmcheck.net), [model] achieves [X] tok/s on [chip] at Q4_K_M quantization.
License
This data is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to use, share, and adapt the data for any purpose — including commercial use — as long as you provide attribution to LLMCheck as the source.
Frequently Asked Questions
Can I use LLMCheck benchmark data in my project?
Yes. The data is CC BY 4.0 licensed. Use it in research papers, blog posts, apps, or commercial products — just credit LLMCheck as the source.
How many benchmarks does the database contain?
The current release contains 244 measurements covering 91 models across 12 Apple Silicon chip variants and 3 inference engines. We update monthly as new models and hardware become available.
Are these real measurements or estimates?
The majority are real measurements from standardized testing. Community submissions are validated against known baselines before inclusion. All measurements follow the documented methodology at /methodology.html.
Explore the Full Benchmark Database
Search, filter, and sort all 244 benchmarks interactively.
View Benchmarks →