# CK, CM and CMX are a legacy automation stack with a named successor

> The mlcommons/ck repository holds Collective Knowledge and its Collective Mind automation frameworks, which turn research projects into file-based repositories of portable artifacts with a shared command line and Python API, and which modularise MLPerf benchmarks. The project says so itself that CK, CM and CMX are legacy and that cMeta is the next generation, and the most recent tagged releases are from March 2025.

**mlcommons/ck** — Collective Knowledge (CK), Collective Mind (CM/CMX) and MLPerf automations: community-driven projects to learn how to run AI, ML, and other emerging workloads more efficiently and cost-effectively across diverse models, datasets, software, and hardware using MLPerf methodology and benchmarks.

- Repository: https://github.com/mlcommons/ck
- Website: https://cTuning.ai/project/cmeta
- Stars: 649 · Forks: 122
- Language: Python
- License: Apache-2.0
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/mlcommons-ck

## The repository names its own successor

The first thing the README says is that CK, CM and CMX are legacy and maintained by MLCommons, and that their successor, cMeta, is being developed by the original author, Grigori Fursin, in a separate repository. That statement is repeated twice, in a callout at the top and again in the body, which tells you how deliberate the positioning is. The legacy status also has a history attached to it. The legacy CM API and CLI, described as the 2021 to 2024 generation, were donated to MLCommons by Grigori Fursin, the cTuning foundation and OctoML, and they are now supported by the MLCommons Infra WG through MLCFlow, the MLC scripts and mlcr. The licence is Apache 2.0, with copyright held by MLCommons for 2021 to 2025 and by the cTuning foundation for 2014 to 2021. In short, this is a donated project under institutional care, and the original author has moved on to the next thing.

## CM turns a project into a repository of artifacts

Collective Mind is the core idea, and it is worth describing precisely because the naming makes it opaque. CM was developed to support open science and to facilitate collaborative, reproducible and reusable research, development and experimentation based on the FAIR principles. Its mechanism is to convert software projects non-intrusively into file-based repositories of portable and reusable artifacts, which means code, data, models and scripts, with extensible metadata. On top of that it offers a unified command line interface and a simple Python API. Because the artifacts are files with metadata, they can be chained into portable, technology-agnostic automation workflows, which is what lets a user rerun and reuse a complex experimental setup across different models, datasets, software and hardware. The README's own example of what this buys you is MLPerf: CM helps to modularize, automate and customize MLPerf benchmarks.

## CMX is the same shape with a simpler interface

CMX, described as Collective Mind eXtension or Common Metadata eXchange, is presented as the next evolution of the Collective Mind framework, designed to enhance simplicity, flexibility and extensibility based on user feedback. The compatibility story is the important part for an adopter. CMX is backwards compatible with CM, released along with CM in the `cmind` package on PyPI, and it can serve as a drop-in replacement for CM and for the legacy MLPerf automations while providing a simpler interface. The newer tooling also comes with its own automation set, CMX4MLOps. The dependency picture is small, which matters for a framework that ends up inside CI images and research environments. The requirements file lists five packages:

```yaml
cmind>=3.5.2
pyyaml
requests
setuptools
giturlparse
```

The framework itself requires Python 3.8 or newer, and the project says the automations run natively on Ubuntu, MacOS, Windows, RHEL, Debian and Amazon Linux, in a cloud or inside automatically generated containers, while keeping backward compatibility.

## script and cache are the two automations worth learning first

The MLOps and MLPerf automations are described as portable, extensible and technology-agnostic recipes with a common command line and Python API, called CM scripts, that unify the manual steps required to compose, run, benchmark and optimize complex ML and AI applications on diverse platforms with any software and hardware. Two of them are singled out as the key automations, `script` and `cache`, and both are browsable in an online catalog on the CK playground and in the online MLCommons catalog. The `script` automation extends the idea of `cmake` with simple Python automations, native scripts and JSON or YAML meta descriptions, which is the mental model to carry into a new workflow: you describe what to build, and the framework works out how to run it. The `cache` automation is the counterpart that keeps results around so a rerun does not redo the work. There is also an artifact evaluation automation, a community-driven initiative that uses CK, CM and CMX to automate artifact evaluation at ML and systems conferences.

## Running MLPerf inference means swapping four command names

For teams that want MLPerf inference benchmarks across diverse systems, the README points at the online MLCommons documentation and at CMX. The installation is a single pip command, `pip install cmx4mlperf`, and then the legacy commands are substituted one for one. The mapping is mechanical:

```bash
cm -> cmx
mlc -> cmlc
mlcr -> cmlcr
-v -> --v
```

Four substitutions, no other changes, and the same benchmark runs under the newer interface. That is a cheap migration path for an existing MLPerf setup, and it is also the clearest sign of how the project is designed: the command line is the stable contract, and the framework behind it is allowed to move. The repository carries the pieces side by side, with `cm/`, `cmx/`, `cm-mlops/`, `cm4mlops/` and `cmx4mlops/` directories, plus `cmx4mlperf/`, a `flexbench/` directory, `main.py`, `cmr.yaml`, a `citation.bib`, an `mkdocs.yml` for the documentation and a `tests/` directory.

## The Playground is an index of automations, not a store of results

The Collective Knowledge Playground is described as a unified and open-source platform designed to index all CM and CMX automations, similar to PyPI, and to help users prepare CM and CMX commands. The three jobs it names are aggregating, processing, visualising and comparing MLPerf benchmarking results for AI and ML systems, running MLPerf benchmarks, and organising open and reproducible optimization challenges and tournaments. The comparison with PyPI is the useful part for an engineer, because it tells you what to expect from the model. Scripts are shared, versioned and indexed, and the playground is the place where you find out which automation already solves the problem you were about to write. The older CM-MLOps project from 2021 and CM4MLOps from 2022 to 2024 are also in the repository as legacy entries, alongside CK automation framework versions one and two under `ck/`.

## Apache 2.0, March 2025 releases, and a push in October 2026

Maintenance is the part to weigh. The tagged releases are CMX v4.1.0 on 2025-02-17, described as fixing legacy interfaces, cmind v4.1.3 on 2025-02-20 and CMX v4.1.4 on 2025-03-13, all three of them bug fixes. The last push to the repository was on 2026-10-20 and the repository is not archived, so work continues on the default branch, but the release line stopped in March 2025, roughly a year and a half ago. Since the documentation also states that the framework is continuously extended by the community and MLCommons members, that gap is worth asking about rather than assuming. The licence is Apache 2.0 with the file in the repository root, and the copyright is split between MLCommons and the cTuning foundation, which is the usual shape for a donated project and means there is no vendor who can be asked to add a feature on a deadline.

## Conclusion

Teams that already run CM or CMX automations, or that want to reproduce MLPerf runs through a script and cache model rather than a bespoke harness, will find a mature body of work here, with MLCommons maintaining it and the artifacts on a public index. It is the wrong starting point for a new project, because the repository itself points you at cMeta for that, and it is the wrong choice if you need a small, dependency-light tool, since the framework is a large surface with a command line, a Python API and a metadata format. Before adopting anything here, check whether your automations are already in the CM format, and read the migration notes for the `cm` to `cmx` command changes before you write scripts against them.

## FAQ

### What is the difference between CM and CMX in the MLCommons CK repository?

CM is the Collective Mind automation framework that turns projects into file-based repositories of portable artifacts with a unified command line and Python API. CMX is the next evolution of it, backwards compatible with CM, released in the cmind package, and able to serve as a drop-in replacement with a simpler interface.

### How do I run MLPerf inference benchmarks with CMX?

The documentation says to install cmx4mlperf with pip and then substitute the legacy commands and flags: `cm` becomes `cmx`, `mlc` becomes `cmlc`, `mlcr` becomes `cmlcr`, and `-v` becomes `--v`. The rest of the run is unchanged.

### Is CK, CM and CMX still being developed?

The README describes CK, CM and CMX as legacy and maintained by MLCommons, and says the next generation, cMeta, is being developed by the original author in a separate repository. The legacy automations are supported by the MLCommons Infra WG, and the most recent tagged releases are from March 2025.

### What are the key CK automations?

The README names two key automations, `script` and `cache`, which are browsable in the online CK playground catalog and the online MLCommons catalog. CM scripts extend the idea of cmake with simple Python automations, native scripts and JSON or YAML meta descriptions.

## Sources

- [License: Apache-2.0](https://github.com/mlcommons/ck/blob/master/LICENSE)
- [mlcommons/ck on GitHub](https://github.com/mlcommons/ck)
- [Project website](https://cTuning.ai/project/cmeta)
- [README](https://github.com/mlcommons/ck/blob/master/README.md)
- [Releases](https://github.com/mlcommons/ck/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/mlcommons-ck