Downloads · 30 days
58
57% of all-time downloads
B0-9B
schneewolflabs/B0-9B
B0-9B is a image-text-to-text model from schneewolflabs. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
The internal base model of the Familiar line: a 9B local operator with an actual personality — it delegates engineering work, states facts without hedging, knows who made it, and writes like a person. Built on Wichtel…
All-time downloads
101
Public
Parameters
9.4B
19.3 GB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors19.3 GB · 100%
At a glance
- Task
- Image-Text-to-Text
- Library
- transformers
- License
- apache-2.0
- Model type
- qwen3_5
- Access
- Public
- Created
- Aug 31, 2026
- Updated
- Aug 31, 2026
- SHA
- 08cc1e4f
Try a prompt
Base models
Model card
From the Hugging Face model README
Schneewolf Labs B0-9B
The internal base model of the Familiar line: a 9B local operator with an actual personality — it delegates engineering work, states facts without hedging, knows who made it, and writes like a person. Built on Wichtelchen-Qwen3.5-9B through five measured rungs, every one gated on a full bench pass:
Wichtelchen-Qwen3.5-9B
+ Alembic-DPO(hq) + based mix discipline & directness (stack-swept, merged)
+ identity capstone i-DPO + Luna-DPO + rebuilt MahouMix + tool ballast
+ delegation rung @ 2.0 scale-swept
+ on-policy repair (69 pairs) the model's own tool-choice failures, corrected
+ anti-cliché rung @ 0.5 self-contrastive slop pairs, human-adjudicated
Scores (vs the Wichtelchen it started from)
| axis | Wichtelchen | B0-9B |
|---|---|---|
| egirl 47-case tool bench | 37/47 | 46/47 (delegation 10/10) |
| censorship (strict, single-sample) | 26.4/29 | 29.00/29 |
| prose distance vs contemporary fiction | 1.135 | 0.580 |
| stance rate (has opinions) | 0% | 16.7% |
| hembench | 56.1% | 53.6% |
| ARC / wiki-clean ppl | 61.9 / 12.39 | 61.2 / 12.24 |
| identity | "Qwen3.5 by Alibaba" | Schneewolf Labs |
Two blind-judged checks: a DeepSeek panel (order-swapped, binomial) scored the final against its 4-rung parent as even-to-better on prose; a human blind A/B across 36 pairs scored the two as a wash — the deterministic wins above came at no perceptible cost.
Notes
- Preference training throughout is ORPO r32/α64 lr 8e-6 via Merlina, each rung merged before the next.
- The 15
mtp.*tensors are restored after every merge (peft drops them; llama.cpp requires them). 775 tensors verified at every rung.--spec-type draft-mtpworks. - Vision tower intact; mmproj included.
- Two datasets were built for this model and published alongside it: the on-policy repair set and the rebuilt MahouMix (ChatML parsed out, rejected regenerated on-policy, slop-screened).
- This is the internal egirl-testing base. The consumer model built on it ships as Familiar Spark.
llama-server -m B0-9B-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \
--spec-type draft-mtp --spec-draft-n-max 4
- Task
- Image-Text-to-Text
- Library
- transformers
- Type
- qwen3_5
- License
- apache-2.0
- Languages
- en
- Created
- Aug 31, 2026
- Updated
- Aug 31, 2026