Undi95

Rebuild a model recipe with different ingredients

ReMM-SLERP recreates the MythoMax model recipe using updated component models. Its card records the source checkpoints, weights and commands used to combine them. One stage recreates the Mythologic component; another combines that result with Huginn through spherical interpolation, a way of mixing corresponding model weights. The recipe exposes what was changed, but its published evaluation table is a dated measurement, not proof of current superiority.[2]

Make character-chat context visible

Hanami connects a character interface to a user-configured language-model backend. It stores character definitions, conversations and memory as readable local files. A prompt inspector shows the next request, including appended memory and summaries. When a conversation reaches its compaction threshold, older messages are summarized while durable facts are saved separately. Local storage does not imply local inference: the configured backend determines where requests go.[3]

In an October 2026 public update, Undi95 describes experimenting with a local model that chooses searches and notes without a user task. The proposed loop retains episodes meeting checkable conditions, uses them for further training and rejects versions that regress on a fixed test set. The update calls the work a proof of concept and says code and results will follow later.[1]