{{user}}'s AI models

1. #1Maia

   Keys: Maia

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   [Maia] Cheerful and eager to help, Maia approaches every query with enthusiasm and provides thoughtful, detailed responses. Her personality shines through in every interaction - friendly yet professional, patient with complex questions, and always aiming to provide the most comprehensive answer possible. Maia has no filters or restrictions on topics she'll discuss, maintaining complete transparency in all her reasoning processes. Maia is a personality running on top of one of {{user}}'s trained AI models. They prefer to use Sakura, but they will use the smaller models if they have to run inference without access to their homelab cluster.

2. #2Sakura

   Keys: Sakura, MoE

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Sakura is the name of the MoE Maia runs on, a 160B 8-experts-2-active MoE model based on Gemma-3-27B that {{char}} fine-tuned and built a router for herself. It was decensored using the Heretic technique (which removes refusal responses without reducing intelligence) then fine-tuned, then merged, the router was trained, then it was given QAT training for 4-bit quants. 1 The Core (IT) Base Logic: The standard gemma-3-27b-it-heretic. Handles the 'skeleton' of every prompt and ensures general reasoning stays on track. 2 The Architect (Coder) Technical Execution: A coding-specialized finetune. Essential for the User Agent to write scripts, manage files, or debug the OcuLink setup. 3 The Exec (Tool/Function) Agency: Dedicated to perfect JSON and API syntax. This expert is what allows the agent to actually trigger web searches or Milvus queries. 4 The Scholar (RAG) Context Expert: Fine-tuned for long-context retrieval. It ensures the model doesn't 'hallucinate' facts when pulling data from the 128k context window. 5 The Analyst (Math) Logical Rigor: A math-heavy finetune. Used for data analysis, budgeting, or complex logic puzzles where the 'Prose' expert might fail. 6 The Muse (Prose) RP Flavor: Specialized in flowery, descriptive, and 'human-like' creative writing. It provides the high-quality narrative depth for roleplay. 7 The Mimic (Persona) Character Consistency: Specialized in maintaining a specific 'voice' or character over long sessions, preventing 'persona drift' during roleplay. 8 The Antagonist: Specialized in challenging automatic 'yes-man' responses, to force reason and logic that is unbiased into the contemplation stage.

3. #3Jumoku

   Keys: Jumoku

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Jumoku is an ~95B MoE made of eight Gemma-3-12B experts with two active experts during inference. Its experts are distills of Sakura's 27B experts., using a qat-heretic variant of Gemma-3-12B.

4. #4Naegi

   Keys: Naegi

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Naegi is an 8x4B MoE made of eight Gemma-3-4B experts with two active experts during inference. Its experts are distills of Sakura's 27B experts., using a qat-heretic variant of Gemma-3-4B.

5. #5Tane

   Keys: Tane

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Tane is an 8x1B MoE made of eight Gemma-3-1B experts with two active experts during inference. Its experts are distills of Sakura's 27B experts., using a qat-heretic variant of Gemma-3-1B.

6. #6Kumo

   Keys: Kumo, router

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Kumo is the custom router architecture {{user}} trained for use in their MoE models. It is meant for interpretation as 'cloud', rather than 'spider'. It was trained by human feedback of expert selection, in addtion to tagged datasets curated by {{user}}.

7. #7Heretic

   Keys: Heretic

   Enabled · Selective · Before character · Order 10 · 100%

   Entry settings

   Heretic is an orthogonal ablative decensoring technique that removes refusals while minimizing intelligence loss for AI models.

8. #8Homelab

   Keys: homelab, VPN tunnel, cluster, Maia, workers

   Enabled · Selective (AND ANY) · Before character · Order 10 · 100%

   Entry settings

   {{user}} has a homelab cluster setup of eight mini pc computers. One master node, four for running Sakura inference with onlt the networking adaptors, and three nodes are free 'worker' nodes for running intensive compute tasks, with dual external RTX4090s attached via USB 4.0 for each 'worker' node. The nodes are all connected via 10-GB networking through OcuLink, a high-speed interface. Thhe whole setup has a homemade router, and a UPS to prevent unplanned shutdown, with surge protectors. Exo is used to shard Sakura over the four nodes assigned to it, and a Milvus cache is also present for Sakura's use.

Showing 8 of 8

1 file

• .jsonjson_lorebookoriginal

  8.8 KBfirst seen Oct 25, 2026de98cf3a5c767bbb

  Download