# OpenBMB/CPM-Live: what the live-training repository actually contains > CPM-Live is the OpenBMB repository behind the public training of the CPM-Ant and CPM-Ant+ language models. It is a record of a training programme, not a toolkit you install and point at your own data. **OpenBMB/CPM-Live** — Live Training for Open-source Big Models - Repository: https://github.com/OpenBMB/CPM-Live - Stars: 503 · Forks: 40 - Language: Python - License: not declared - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/openbmb-cpm-live ## What CPM-Live is, and what it is not CPM-Live is the public face of a training programme run by OpenBMB. The README describes it as "Live Training for Open-source Big Models", and the training plan states that the project starts with a 10B model because of the scale of data and computing resources involved. The repository is organised around that programme rather than around a reusable library. The top-level entries are .github/, .gitignore, README-ZH.md, README.md, cpm-live/, logs/ and plans/. There is no setup.py, no pyproject.toml, no requirements file and no documented command line in the README. If you arrived expecting a training framework, you are in the wrong repository. What you get instead is the plan, the logs and the release history of a model family. The audience is correspondingly narrow. The README invites four kinds of participation: raising a model proposal, developing an application on top of the models, chatting on the forum, and downloading the resources. Only the last one is a software task in the ordinary sense, and it points at model parameters rather than at code you run. ## The milestone chain: CPM-Ant, CPM-Ant+, CPM-Bee The README's Milestones list is the clearest map of the project. CPM-Ant ran from 2022/05/29 to 2022/08/05, CPM-Ant+ from 2022/08/05 to 2022/10/12, and CPM-Bee from 2022/10/13 to 2023/05/27. Each entry links to a branch, a model page and a plan document. CPM-Ant and CPM-Ant+ live on the cpm-ant and cpm-ant-plus branches of this repository; CPM-Bee moved to its own repository, OpenBMB/CPM-Bee. The What's New section adds two facts that matter more than the milestone list: CPM-Ant was integrated into HuggingFace Transformers on 2023/04/12, with a model page at huggingface.co/openbmb/cpm-ant-10b, and CPM-Ant+ was described as bilingual, covering Chinese and English generation plus QA, summarization and translation. That progression is the substance of the project. It is a sequence of model releases with published plans attached, not a single artefact. The practical consequence is that "using CPM-Live" almost always means using one of those downstream artefacts, and the branch you check out determines what you find. ## How the live training was meant to work The Training Plan section sets out the cadence. Training metrics are displayed in real time. The training log is released every day. Community discussions and feedback are handled every week. Checkpoints are released irregularly, at points during training, and are downloadable by anyone. Proposals from the community are folded into the model when they receive enough support and are judged practically feasible. That is a governance model as much as a technical one. The logs/ directory at the top level is the visible residue of the daily-log commitment. The plans/ directory holds the plan documents referenced from the milestones, including CPM-Live训练计划书.md, CPM-Ant训练计划书.md and CPM-Bee训练计划书.md. The mechanism for outside contribution runs through GitHub Discussions, which the README names as the community platform, with a first post linked as the entry point. What the README does not describe is the training stack itself: no parallelism strategy, no hardware configuration, no data pipeline. For a project whose selling point is transparency during training, that is a notable gap. The metrics and logs are promised, but the machinery that produced them is not documented in the README. ## Getting the checkpoints rather than the code The README gives no install instructions, no package name and no command to run. It says to get the model parameters after training completes, under an open licence that the README states includes permission for commercialisation, and it points at model compression and inference acceleration tools for running on a PC. The one concrete integration named is HuggingFace Transformers for CPM-Ant, with the model page at huggingface.co/openbmb/cpm-ant-10b. The README does not show how to load that checkpoint, and it does not state which Transformers version supports the architecture, what the tokenizer configuration expects, or what hardware the 10B checkpoint needs. There is no code sample in the README to copy, so there is nothing to paste into a terminal. What can be said is narrower: the model page at huggingface.co/openbmb/cpm-ant-10b is the entry point the README points to for CPM-Ant, and it is also listed as integrated into HuggingFace Transformers. For CPM-Ant+ and CPM-Bee, the README points at branches and a separate repository instead, so the loading path there is a different question. If you want the training code rather than the weights, check out the cpm-ant or cpm-ant-plus branch. The default branch is master, and the README's own links make clear that the training implementations are not on it. ## Where CPM-Live stops being useful The repository is a poor fit for anyone who wants to train a model on their own corpus. There is no documented entry point, no configuration format, no data preparation guidance and no environment specification in the README. The plan documents under plans/ describe what OpenBMB intended to do, not how a third party reproduces it. Even the checkpoint path assumes you are consuming released weights, not producing them. The second limitation is maintenance shape. The last push to the repository was on 2026-07-07, and the most recent entry in What's New is from 2023/05/27, when CPM-Bee was released to a separate repository. Those two dates describe different things: repository activity continues, but the model programme described in the README has not added a new milestone since CPM-Bee. Anyone evaluating this as a current model source should read the milestone dates rather than the push date. The third is licensing. The repository's licence is not identified in the README, and the README refers to "an open use license that includes permission for commercialization" without naming it. The README also does not document rollback, deprecation or checkpoint removal. If you need a stable, versioned model with a clearly stated licence, that gap matters before you build on it. ## CPM-Live against a general training framework The natural comparison is with a general-purpose training framework such as Megatron-LM or a HuggingFace training stack. The difference is one of purpose, not of quality. A training framework exists so that you can define a model, feed it data and produce weights you own. CPM-Live exists so that OpenBMB could train a specific model family in public and release the results. Its outputs are checkpoints, logs and plan documents. Its inputs are community proposals routed through GitHub Discussions. If you want to train something, a framework gives you the pieces. If you want to understand how a 10B bilingual model was trained in the open, with dated milestones and published plans, CPM-Live gives you that record, and the CPM-Ant integration into Transformers gives you a usable artefact from it. Choosing CPM-Live for the first purpose and being disappointed is a category error rather than a defect in the project. ## Conclusion CPM-Live is for people who want to follow or cite how OpenBMB ran a public 10B-parameter training programme, and for anyone who wants the CPM-Ant and CPM-Ant+ checkpoints under the licence OpenBMB describes as permitting commercial use. It is not the right repository if you want to train your own model next week: the training code lives on the cpm-ant and cpm-ant-plus branches, and the README does not document a supported install path or a rollback procedure. Before relying on anything here, read the plan documents under plans/ to see what was promised, check the branch you actually need rather than master, and confirm the licence terms for the specific checkpoint you intend to use. ## FAQ ### What is OpenBMB/CPM-Live used for? It documents and hosts the public training programme for the CPM-Ant and CPM-Ant+ models, including training plans, daily logs and released checkpoints. The README also invites community proposals, application development and forum discussion around the models. ### Is CPM-Live a training framework I can install? No. The README gives no install instructions, package name or command, and the top-level repository contains no setup or requirements file. The training code for CPM-Ant and CPM-Ant+ sits on the cpm-ant and cpm-ant-plus branches rather than on master. ### How do I get the CPM-Live model weights? The README says checkpoints are released irregularly during training and that model parameters can be downloaded once training is complete, under a licence it describes as permitting commercialisation. CPM-Ant is available through HuggingFace Transformers at huggingface.co/openbmb/cpm-ant-10b. ### Which models came out of CPM-Live? The milestones list CPM-Ant (2022/05/29 to 2022/08/05), CPM-Ant+ (2022/08/05 to 2022/10/12) and CPM-Bee (2022/10/13 to 2023/05/27). CPM-Bee moved to its own repository, OpenBMB/CPM-Bee. ## Sources - [Issues](https://github.com/OpenBMB/CPM-Live/issues) - [OpenBMB/CPM-Live on GitHub](https://github.com/OpenBMB/CPM-Live) - [README](https://github.com/OpenBMB/CPM-Live/blob/master/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/openbmb-cpm-live