B0-27B

Schneewolf Labs B0-27B

The internal base model of the Familiar line at 27B: a local operator with an actual personality that completes real engineering work end-to-end. Built on Hemlock-Qwen3.8-27B (WichtelHui-SLERP + the Hemlock coding adapter — a base that starts at 47/47 on our operator bench) through two measured rungs, every step gated on a full bench pass:

Hemlock-Qwen3.8-27B
 + based mix @ 1.0 (merged)        directness substrate — weasel + grok-pi + seX-ai
 + identity capstone (ORPO 2ep)    i-DPO + Luna-DPO + on-policy MahouMix rebuild
                                   + tool ballast + destructo — trained ON the merge,
                                   so the capstone overwrites the substrate's costs
                                   and keeps its gains (the seesaw, used on purpose)

Scores (vs the Hemlock base it started from)

The egirl-bench dip is five known single-tool-call reflexes (three delegation prompts answered with git_status, two trivial edits answered with a file search). They do not reach task completion: on kirabench — six sandboxed engineering tasks driven through the real egirl harness with Codex as the code agent, judged only by deterministic checkers — B0-27B goes 6/6 and fixed a failing-test task ~9× faster than its base.

Notes

• Preference training via Merlina (ORPO r32/α64 lr 8e-6); the substrate adapter was hand-merged into the weights (delta surgery — vision and MTP tensors byte-identical to the base) before the capstone trained on the result. 1,199 tensors verified at every merge.
• MahouMix was rebuilt on-policy: ChatML parsed out, rejected regenerated from the substrate-merged model itself.
• Vision tower and the 15 mtp.* tensors are untouched from the base; mmproj included in the GGUF repo. --spec-type draft-mtp works.
• This is the internal egirl-testing base. The consumer model built on it ships as Familiar Ember.

llama-server -m B0-27B-Q8_0.gguf -ngl 99 -c 16384 --jinja -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 4