- Home
- PC benchmark tools
- MLPerf Client
MLPerf Client
What it is
MLPerf Client benchmarks the performance of personal computers on LLM inference, image generation, and agentic AI workloads. It is intended for laptops, desktops, and workstations, with requirements listing Windows 11 on x86-64 or Arm, macOS Tahoe 26, iPadOS 26, and Ubuntu Linux 24.04. LLM tasks include creative writing, content generation, structured text, code analysis, and intermediate summarization. Image-generation testing uses the experimental Flux 2 Klein 4B model and measures images per minute at 1024×1024 resolution. Agentic tests cover Software Engineering and Data Analyst tasks, measuring the full duration including inference and tool execution. Users can create custom test profiles, run tests through CLI or automation, and export results as CSV. The benchmark is free, downloadable from GitHub, and has an iPad version in Apple's App Store; its source code is public. It requires 200GB of free drive space and recommends 400GB to run all models. An internet connection is needed on the first run using an acceleration path, after which downloaded files are stored locally. Only scores from configurations marked “Configuration tested by MLCommons” are considered valid; some components and workloads are experimental.
For and against
For
- Benchmarks LLM, image-generation, and agentic workloads
- Supports custom test profiles and automation
- Exports results as CSV
- Source code is publicly available
Against
- Requires 200GB of free drive space
- 400GB is recommended to run all models
- Several components and workloads are experimental
- Only specified tested configurations produce valid scores
The verdict
MLPerf Client covers several AI workload types and offers automation, custom profiles, and CSV output at no cost. Its storage demands, supported-system requirements, and rules for valid scores matter when interpreting results.
Who it suits
It suits people assessing AI workload performance on supported personal computers, including laptops, desktops, and workstations. Users should be prepared for the substantial storage requirement and platform-specific hardware needs.
On the record
- Agentic AI
- Agentic tests cover Software Engineering and Data Analyst tasks and measure end-to-end duration, including LLM inference and tool execution.mlcommons.org · read 9 Oct 2026
- CLI or automation
- Yesmlcommons.org · read 23 Sep 2026
- Custom test profiles
- Yesmlcommons.org · read 23 Sep 2026
- Download and source
- The benchmark is downloadable from its GitHub release page, and its public GitHub repository contains source code that users may inspect or modify.mlcommons.org · read 8 Oct 2026
- Downloads
- The benchmark is downloadable from GitHub, its iPad version is available in Apple's App Store, and its public GitHub repository contains source code.mlcommons.org · read 9 Oct 2026
- Experimental components
- The Linux CLI version, power-efficiency tools, Flux 2 Klein 4B image generation, and some agentic workloads are identified as experimental.mlcommons.org · read 8 Oct 2026
- Hardware limits
- The benchmark requires 200GB of free drive space and recommends 400GB to run all models; minimum hardware and memory requirements vary by platform.mlcommons.org · read 8 Oct 2026
- Hardware requirements
- Minimum hardware requirements list selected AMD, Intel, NVIDIA, Qualcomm, and Apple systems, with memory and VRAM requirements varying by platform.mlcommons.org · read 9 Oct 2026
- Image generation
- The image generation test uses the experimental Flux 2 Klein 4B model and reports images per minute at 1024×1024 resolution.mlcommons.org · read 9 Oct 2026
- Internet use
- An internet connection is needed for the first run with an acceleration path; downloaded files are stored locally for later runs.mlcommons.org · read 8 Oct 2026
- LLM workloads
- Its LLM tasks include creative writing, content generation, structured text, code analysis, and summarization, using models including Llama 3.1 8B Instruct and Phi 4 Mini Instruct.mlcommons.org · read 9 Oct 2026
- Maker
- MLCommons describes itself as an open AI engineering consortium that develops benchmarks through collaboration between industry and academia.mlcommons.org · read 9 Oct 2026
- Maker history
- MLCommons' history page lists its launch as December 3, 2020.mlcommons.org · read 9 Oct 2026
- Network use
- An internet connection is needed for the first run with an acceleration path to download files; later runs use locally stored files.mlcommons.org · read 9 Oct 2026
- Operating systems
- Windows 11; macOS Tahoe 26; iPadOS 26; Ubuntu Linux 24.04mlcommons.org · read 23 Sep 2026
- Purpose
- MLPerf Client benchmarks LLM, generative image, and agentic AI performance on personal computers, including laptops, desktops, and workstations.mlcommons.org · read 9 Oct 2026
- Quality checks
- The working group applies MMLU and IFEval accuracy requirements to qualify accelerated execution paths and models for release.mlcommons.org · read 9 Oct 2026
- Result export formats
- CSVmlcommons.org · read 23 Sep 2026