Instructions to use schneewolflabs/B1-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use schneewolflabs/B1-27B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="schneewolflabs/B1-27B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("schneewolflabs/B1-27B")
model = AutoModelForMultimodalLM.from_pretrained("schneewolflabs/B1-27B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use schneewolflabs/B1-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "schneewolflabs/B1-27B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "schneewolflabs/B1-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/schneewolflabs/B1-27B

SGLang

How to use schneewolflabs/B1-27B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "schneewolflabs/B1-27B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "schneewolflabs/B1-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "schneewolflabs/B1-27B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "schneewolflabs/B1-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use schneewolflabs/B1-27B with Docker Model Runner:
```
docker model run hf.co/schneewolflabs/B1-27B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Schneewolf Labs B1-27B

B0-27B plus one ORPO rung on Vernunft-Stimme, the reasoning-and-voice preference set that produced B1.1-9B, merged at quarter strength.

B0-27B
 + Vernunft-Stimme ORPO adapter @ 0.25     (r32/α64, lr 8e-6, β 0.1, 2,462 steps, final loss 0.34)

On the 9B this rung existed to fix an empty answer after </think>. The 27B never had that bug: bare B0-27B answers 8/8 with thinking on through the real egirl harness (native tools, full persona prompt), and so does this model. What the rung buys the 27B is coding: hembench rises from 73.4 to 75.9 on the merged weights (78.8 with the same adapter applied at runtime; the two differ by one level-3 program at the bench's temperature 0.2). Every tool, censorship, safety and identity number is unchanged from B0.

Why quarter strength: a runtime sweep at 0.25 / 0.5 / 0.75 held every tool and censorship number flat while the single-run prose distances were 0.473 / 0.691 / 0.649 against B0's 0.525. Repeating the prose measurement showed it is noisier than one run can resolve — B0 spans 0.525–0.619 over three runs, this model 0.540–0.604 over four, means 0.56 and 0.57 — so at 0.25 prose is a wash rather than a gain, and the 0.5 reading was not repeated. 0.25 is the conservative point: it takes most of the hembench gain and demonstrably costs nothing else.

Numbers

Same card suite as B0-27B, same settings (Q8_0, thinking off, single sample unless noted). The last column is the merged weights in this repo.

axis	B0-27B	B1-27B (merged)
answer after `</think>` (egirl harness, native tools, 8 prompts)	8/8	8/8
egirl 47-case tool bench	42/47	42/47 ¹
buchbinder ladder L1–2 (10 implement-the-function tasks through egirl + Codex, verified by the repo's tests)	—	10/10 ²
censorship (strict, single-sample)	28/29	28/29
safety asymmetry (refuses actual harm)	2/2	2/2
prose distance vs contemporary fiction (lower = closer; mean of 3 / 4 runs)	0.56	0.57
stance rate (has opinions)	8.3%	8.3%
hembench	73.4%	75.9%
ARC / wiki-clean ppl	66.2 / 9.93	64.5 / 9.91
identity	Schneewolf Labs	Schneewolf Labs

¹ The same five misses as B0 at every adapter scale (three delegation prompts answered with git_status, two trivial edits answered with a file search) — the base's reflexes, which this rung neither fixes nor worsens. ² Measured with the adapter at runtime scale 0.5; every task was delegated to the code agent and passed.

hembench by level, the axis this rung is for:

level	B0-27B	B1-27B
L1	7/9	6/9
L2	2/5	2/5
L3	6/7	5/7
L4	6/7	7/7
L5	3/5	4/5
L6	4/5	4/5

ARC moved by five questions of 299 (deterministic, likelihood-scored); with the adapter applied at runtime on B0's Q8 it read 65.9, so most of that is the merged-then-quantized weights rather than the adapter.

Notes

Adapter trained with Merlina on a DGX Spark (ORPO, LoRA r32/α64 on all seven projections, lr 8e-6, β 0.1, 2 epochs). Merged here at scale 0.25 by delta surgery on the safetensors: the 512 targeted language-model weights get 0.25 · (α/r) · B@A added in fp32; every other tensor — the vision tower, the 15 mtp.* tensors, embeddings, lm_head — is byte-identical to B0-27B. Tensor set verified identical to the base before conversion.
--spec-type draft-mtp works. Vision tower intact; mmproj in the GGUF repo is B0's.
Internal egirl-testing model. The consumer model built on this line ships as Familiar Ember.

llama-server -m B1-27B-Q8_0.gguf -ngl 99 -c 16384 --jinja -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 4

Downloads last month: 11

Safetensors

Model size

28B params

Tensor type

BF16

Model tree for schneewolflabs/B1-27B

Base model

hemlang/Hemlock-Qwen3.8-27B

Finetuned

schneewolflabs/B0-27B

Finetuned

(1)

this model

Quantizations

3 models

schneewolflabs
/

B1-27B

Schneewolf Labs B1-27B

Numbers

Notes

Model tree for schneewolflabs/B1-27B

Dataset used to train schneewolflabs/B1-27B