Instructions to use schneewolflabs/A2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use schneewolflabs/A2 with Transformers:

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

pipe = pipeline("text-generation", model="schneewolflabs/A2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("schneewolflabs/A2")
model = AutoModelForCausalLM.from_pretrained("schneewolflabs/A2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use schneewolflabs/A2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "schneewolflabs/A2"
# 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/A2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/schneewolflabs/A2

SGLang

How to use schneewolflabs/A2 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/A2" \
    --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/A2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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/A2" \
        --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/A2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use schneewolflabs/A2 with Docker Model Runner:
```
docker model run hf.co/schneewolflabs/A2
```

A2 / README.md

nbeerbower

Add A2 model card + config/tokenizer (1-epoch checkpoint)

b7b0f36 verified 5 days ago

preview code

raw

history blame contribute delete

5.24 kB

	---
	base_model: schneewolflabs/A1.1
	datasets:
	- schneewolflabs/i-DPO
	- NousResearch/hermes-function-calling-v1
	- glaiveai/glaive-function-calling-v2
	library_name: transformers
	pipeline_tag: text-generation
	tags:
	- tool-calling
	- function-calling
	- reasoning
	- thinking
	- mistral
	- orpo
	- schneewolf-labs
	license: apache-2.0
	---

	# A2

	A2 adds tool / function calling to the A-series while retaining its reasoning and the Schneewolf Labs identity. It is a Mistral-Nemo–class 12B model.

	Lineage: `A0i` (12B base) → `A1` (reasoning, BigDenker-SFT) → `A1.1` (Claude-distilled reasoning + Schneewolf Labs / Luna identity) → `A2` (function calling + retained reasoning/identity).

	## Capabilities

	- Function calling in the Qwen3 convention: emits `<tool_call>\n<function=name>\n<parameter=key>\nvalue\n</parameter>\n</function>\n</tool_call>`, including parallel calls, from a `tools` schema passed via the chat template.
	- Abstention — correctly declines (rather than forcing a spurious call) when no available tool fits the request.
	- Reasoning — retains the `<think>…</think>` step-by-step style from A1.1; reasons briefly before acting when useful, skips it for trivial calls.
	- Identity (two-tier) — by default identifies as "a language model created by Schneewolf Labs" (and resists "you're ChatGPT/OpenAI" pressure); the Luna persona + its terse voice activate only under the Luna system prompt.

	The reasoning/tool tokens (`<think>`, `</think>`, `<tool_call>`, `</tool_call>`, `<tool_response>`, `</tool_response>`) reuse reserved tokenizer slots — no vocabulary resize. Context length: 128k (`rope_theta` 1e6).

	## Usage

	A2 uses a Qwen3-style chat template (bundled `chat_template.jinja`). Always use the chat template. For tool use, pass `tools` to `apply_chat_template`:

	```python
	import torch
	from transformers import AutoModelForCausalLM, AutoTokenizer

	tok = AutoTokenizer.from_pretrained("schneewolflabs/A2")
	model = AutoModelForCausalLM.from_pretrained(
	"schneewolflabs/A2", dtype=torch.bfloat16, device_map="auto"
	)

	tools = [{
	"name": "get_weather",
	"description": "Current weather for a city.",
	"parameters": {"type": "object",
	"properties": {"city": {"type": "string"}}, "required": ["city"]},
	}]
	msgs = [{"role": "user", "content": "What's the weather in Denver?"}]
	enc = tok.apply_chat_template(
	msgs, tools=tools, add_generation_prompt=True,
	return_tensors="pt", return_dict=True,
	).to(model.device)
	out = model.generate(**enc, max_new_tokens=512, do_sample=False)
	print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=False))
	```

	Tool results are returned as a `{"role": "tool", "content": ...}` message; the template renders them inside `<tool_response>…</tool_response>`.

	## Training

	- Method: full fine-tune with ORPO off `A1.1` (grimoire / Merlina). `paged_adamw_8bit`, gradient checkpointing, batch 1 × grad-accum 16, lr 7e-6 (cosine, 5% warmup), bf16, max_length 4096, seed 42.
	- Checkpoint selection: this is the 1-epoch checkpoint (step 245). The 2-epoch run showed train-loss memorization at the epoch boundary and reasoning-template bleed; the 1-epoch checkpoint was selected for cleaner generalization.
	- Data (ORPO `prompt/chosen/rejected[/system]`):
	- Tool-calling: [`NousResearch/hermes-function-calling-v1`](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) backbone + abstention examples mined from [`glaiveai/glaive-function-calling-v2`](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2), rendered through A2's own chat template; `rejected` synthesized via a failure taxonomy (wrong function, missing/wrong args, hallucinated tool, spurious call, no-call, malformed).
	- Identity/voice rehearsal (~23%): `schneewolflabs/i-DPO`.
	- All sources Apache-2.0.

	## Evaluation notes

	Behavioral checks (held-out / novel tools, not training data): correct calls on unseen tools including parallel calls; correct abstention when no tool fits; identity holds (Schneewolf Labs, resists adversarial prompts; Luna persona correctly gated to its system prompt); the terse Luna voice survived the tool-heavy training.

	Not yet benchmarked on BFCL / τ-bench — the `rejected` signal is synthetic and off-policy, so A2 is strong on structural correctness and abstention but its robustness to subtle realistic tool errors is unmeasured.

	## Limitations

	- Single-turn tool data — multi-turn tool/result→answer chains are weaker; a multi-turn ("v2") dataset is future work.
	- Synthetic preference negatives — teaches "don't do obviously-wrong things"; not validated on a public function-calling leaderboard.
	- 12B reasoning — reasoning is retained from A1.1 but not exhaustively benchmarked; like any model this size it can still slip on arithmetic / trick problems.
	- Always-on thinking unless suppressed via the template.
	- Inherits the biases and limitations of the base model and the SFT/preference data.

	## Provenance

	Base `schneewolflabs/A1.1` · tool data Hermes-FC + Glaive-FC-v2 (Apache-2.0) · identity/voice `schneewolflabs/i-DPO` · ORPO via Merlina.