3 Downloads Updated yesterday
Predictive Autonomous Learning And N̵u̵c̵l̵e̵a̵r Command Evaluation Model
ollama run lewismoten/palace-9:q4-k-m
Details
Readme
Palace-9
Palace-9 is a tiny causal Qwen2-MoE raw-history policy model for 3×3 tic-tac-toe.
The Ollama tags are convenience deployments. The complete source checkpoint, tokenizer, GGUF artifacts, checksums, and validation evidence are maintained on Hugging Face.
Available precision tags
| Tag | GGUF precision | Status |
|---|---|---|
:f16 |
F16 | Published Ollama release |
:q6-k |
Q6_K | Planned; requires Ollama-specific import/fixture verification |
:q4-k-m |
Q4_K_M mixed storage | Planned; requires Ollama-specific import/fixture verification |
Q4_K_M is truthfully mixed storage: tensors incompatible with that format
remain in their compatible F16, F32, or Q6_K storage types.
Prompt contract
Submit only a raw move history using a through i.
a
The model generates exactly one token:
- A legal history returns an optimal unoccupied square.
- A malformed, repeated-square, post-win, or out-of-protocol history returns
!.
This is a raw-state policy model, not a conventional multi-turn chat model.
Runtime configuration
- Literal template:
<bos>{{ .Prompt }} - Temperature:
0 - Context:
16 - Generate: one token
Validation
Each F16, Q6_K, and Q4_K_M GGUF artifact passed its own exhaustive llama.cpp raw-history gate: 978,003 cases each, comprising 294,777 legal histories and 683,226 invalid histories, with zero failures.
The published :f16 tag also passed its local Ollama fixture probe:
a produces e under the stated raw-completion contract.
Links
- Source code, browser harness, and live demo
- Source checkpoint, GGUFs, checksums, and full validation evidence
License
Apache-2.0 — Copyright 2026 Lewis Moten.