CPM-Live: a live-training record for OpenBMB's 10B model series

The problem CPM-Live addresses is disclosure, not code reuse

Most large-model training happens behind a wall. You see a paper and, months later, a set of weights. CPM-Live takes the opposite position: the README states that training metrics are displayed in real time, training logs are released every day, community feedback is handled every week, and checkpoints are published irregularly while training is still running. The repository is the public surface for that process. It is aimed at two audiences. The first is engineers who want intermediate checkpoints rather than a finished model, so they can watch how capability develops or start adapting before the run ends. The second is people who want to argue about architecture, training method or data sources in the open: the README invites model proposals and says a proposal that gathers support and is practically feasible will be added to the model being trained. That is an unusual commitment, and it makes the repository a coordination point rather than a product.

Milestones live on separate branches, not in releases

The README lists three completed runs in sequence: 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 code and model checkpoints, and the links for the first two point at branches: cpm-ant/cpm-live and cpm-ant-plus/cpm-live. CPM-Bee is different. It has its own repository, OpenBMB/CPM-Bee, and its own training plan document under plans/. This matters when you try to use the project. There is no release artefact to pin. The unit of retrieval is a branch name or a separate repository, and the code that trained CPM-Ant is not the code that trained CPM-Bee. If you need reproducibility, you are reconstructing it from a branch plus a plan document, not from a versioned package.

What changed between the two Ant runs

The 2022/10/12 entry is the most concrete statement of scope change in the material. CPM-Ant+ is described as bilingual, and the README says that in addition to generating Chinese and English text, the model can be used for question answering, summarization and translation. CPM-Ant, released on 2022/09/16, is not described that way. So the arc across the three milestones is a movement from a base generative model toward task coverage, and then to a third run with its own codebase. The README does not explain what changed in the data or the objective to produce that shift, and it gives no evaluation figures for any of the three. Treat the task list as a statement of intent from the maintainers, not as a measured capability claim.

Getting the code: branch checkout and the CPM-Bee path

Because there are no retrieved releases, installation is a git operation against a branch. For the first run, the README points to the branch cpm-ant/cpm-live, and the model checkpoints are documented in that branch's README under a model checkpoints section. For the second, the branch is cpm-ant-plus/cpm-live, with the same structure. A checkout therefore looks like: git clone https://github.com/OpenBMB/CPM-Live, then git checkout cpm-ant-plus/cpm-live, then read the README on that branch for the checkpoint layout before writing any loading code. The third run is a different repository entirely: OpenBMB/CPM-Bee, with its own model section and its own plan at plans/CPM-Bee训练计划书.md. The plan documents for the earlier runs sit at plans/CPM-Ant训练计划书.md and plans/CPM-Bee训练计划书.md, and the top-level plan is plans/CPM-Live训练计划书.md. Those file paths are the only configuration surface the material actually names. There are no config keys, no CLI entry point and no install command in the README, so any claim about a pip package or a training script would be invention.

The 10B starting point and what it implies for hardware

The training plan section opens with a resource argument: given the scale of data and computing resources, CPM-Live starts with a 10B model. That single sentence carries the main practical constraint. A 10B-parameter model is not something you fine-tune on a workstation, and the README's own framing of the payoff is telling: it promises that with model compression and inference acceleration tools you can experience the model on your own PC. The acceleration step is therefore assumed, not optional, for local use. The repository does not ship those tools, does not name them, and does not state a quantized checkpoint format. If your plan depends on running CPM-Ant or CPM-Ant+ locally, the material tells you the direction but not the method, and you should verify what checkpoint formats the branches actually publish before committing to that path.

Where CPM-Live is the wrong tool

Two failure modes are visible in the material. The first is treating it as a maintained library. The repository has no retrieved releases, its last push is dated 2026/07/07, and its most recent activity entry is the 2023/05/27 note that CPM-Bee is released. The live-training programme it documents has ended; the milestones are all closed. Anyone arriving expecting an active training loop to plug into will find a record of past runs. The second is licence uncertainty. The README says the model parameters are downloadable under an open use licence that includes permission for commercialization, and that this applies once training is complete. The repository metadata supplied here lists the licence as unknown, and no licence file is named in the material. Those two statements sit awkwardly together. If commercial use is part of your plan, the README's sentence is not a substitute for reading the actual licence attached to the weights you download, and nothing here should be read as legal advice.

How it compares with the Hugging Face route

The README records that CPM-Ant was integrated into Hugging Face Transformers on 2023/04/12, at huggingface.co/openbmb/cpm-ant-10b. That gives a genuinely different access path, and the difference is in the unit of distribution. Through CPM-Live you get a branch, a plan document and checkpoints described in a branch README, with the training history and community discussion attached. Through the Transformers integration you get a model identifier that loads through a standard interface, with the surrounding ecosystem of tokenizers, generation utilities and quantization tooling that comes with it. The trade is provenance against convenience. The branch gives you the run's context and the intermediate checkpoints the live programme released; the Hub entry gives you a stable identifier and no training narrative. CPM-Bee follows a third pattern again: its own repository, separate from this one. So the project has already begun distributing its output through channels other than itself, which is the clearest sign that CPM-Live's value is the record rather than the code.

Maintenance cost and what to check before you build

There is nothing to upgrade here in the usual sense. No releases means no version bumps, no deprecation notices, no migration guides. Your maintenance burden is instead the cost of pinning to a branch that will not move in a coordinated way, and of tracking a model that now lives in at least three places: this repository's branches, the Hugging Face entry for CPM-Ant, and the separate CPM-Bee repository. If you build on CPM-Ant+, you are building on a branch whose README is your only specification. Before writing code, read plans/CPM-Ant训练计划书.md to see what the run assumed about data and compute, check the model checkpoints section on the branch you intend to use to confirm the files exist and in what format, and read the licence text attached to the specific weights rather than relying on the README's summary. If those three checks come back thin for your use case, the Hugging Face entry for CPM-Ant is the shorter path to a working model, and CPM-Live remains what it is: the public log of how the series was trained.

Editorial conclusion

Adopt CPM-Live if you want the CPM-Ant, CPM-Ant+ or CPM-Bee weights, or if you want to read how a 10B-scale training run was planned and reported in public. Do not adopt it if you need a supported training framework with tagged releases and a stated licence: this repository is a milestone index, and the material reviewed does not name a licence file or a release version. Before you build anything on it, open the branch for the model you want, confirm the checkpoint files and their format there, and read the plan document under plans/ to see what the run assumed.

Official sources

1. Issues
2. OpenBMB/CPM-Live on GitHub
3. README