AI and ML tooling notes for 2026-09-23

Two low-volume GitHub threads touched local LLM server APIs. Both are open feature or limitation notes rather than shipped changes.

HTTP slot export and import in llama.cpp

A single opening post in ggml-org/llama.cpp asks maintainers to add HTTP routes for exporting and importing slots on the llama.cpp server. The request is framed as a feature proposal for the existing server component. Readers who run or extend llama.cpp deployments may watch whether the maintainers take up the idea.

Logprobs limited to top tokens in Ollama

A user reports that the Ollama logprobs API only returns probabilities for tokens that already rank inside top_logprobs, and points to raising that cap as the practical path. The thread consists of the report and that redirection. Developers who need exact per-token log probabilities outside the current top set will care about any follow-up change to the limit.