bundle-20261018f, released by professorpalmer/bonsai-ada-surgery on October 19, 2026, improves the save layer for long conversations. In an alternating conversation, the state right after each answer is stored as a delta, a format that records only the part that differs from the previous state. When the new prompt of the next turn contains the last saved token as a prefix, restoration now starts from that point. The changes are presented in the GitHub release notes.
The repository presents Bonsai 2 27B along with a Windows bundle and says the ternary weights stay the same. According to the repository description, it supports a 262k context at 12 GB, with decoding speeds from 54 to 97 tok/s and 2x prefill. The q8_0 KV recipe and 9/9 grammar-based tool calling are also part of the description. This release deals with conversation saving: it stores conversations that share the same starting point separately, and adjusts the retry interval when a failure repeats. Under memory pressure it switches a long delta chain to a full save, and protects the saved state of other conversations that were recently in use. A setting to turn the feature off with an environment variable was added as well.
However, neither the name of the environment variable nor the value of the retry interval is given in the release notes. The speed and context figures in the repository description come with no measurement conditions, and the accuracy of delta restoration is only what the release notes state. This change can be read as an improvement to the way saving and restoration work, but the additional conditions and validation data behind the listed figures and the restoration results cannot be checked.