By NODVIEW editors

Sources

Three Tinker checkpoint types: what to check for Inkling 2

Resuming training, sampling a model and exporting weights are different jobs. Read Tinker's checkpoint documentation before fine-tuning Inkling 2 Small.

Before fine-tuning Inkling 2 Small, decide what you need afterward: resume training, use the model within Tinker or deploy it elsewhere. In its November 5 announcement, Thinking Machines Lab offered training and sampling through Tinker but tied downloadable checkpoints to a future weights release. Apache 2.0 is stated as an intention following external safety work, not proof of an artifact you already possess. The reviewed model card lists Tinker Terms of Use in its license field.

The checkpoint guide describes three operations. save_state preserves training state, including Adam's moment estimates; save_weights_for_sampler saves weights for sampling within the service without gradient or optimizer state; save_weights_external prepares weights for use outside Tinker. Their paths contain weights, sampler_weights and external_weights respectively. A tinker:// address alone does not establish what was retained or whether it can resume training from the same state.

Check publishing separately from downloading. The same guide says publishing a checkpoint lets other Tinker users load it from its path, and only the training-run owner can publish it. This differs from obtaining files for external deployment. For get_external_weights_urls, the documentation requires an external_weights path and returns signed links with an expiry time. A generic method description does not override a particular model's availability conditions.

LoRA adds a dependency on the base model. The merging guide explains how to combine adapter changes with the same base used for training to produce a standalone Hugging Face directory. Its example uses Qwen, so it is not evidence that Inkling 2 export is ready. You need both the saved changes and the correct, obtainable base. This is a prerequisite of the documented method, not a result from our own run.

Before paying for training, record the base name and version, required checkpoint type, export conditions and retention period. Tinker supports a TTL; without one, the documentation says a checkpoint remains until deleted. For a handoff, add the training-run identifier and a way to verify downloaded files once they are actually available. NODVIEW compared the announcement, model card and documentation; we did not train, export or run a model locally. The record helps distinguish a migration plan backed by available artifacts from one dependent on a future capability.