Instructions to use Rumiii/Qwen3-8B-Maxima_A1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Rumiii/Qwen3-8B-Maxima_A1 with Transformers:

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

pipe = pipeline("text-generation", model="Rumiii/Qwen3-8B-Maxima_A1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Rumiii/Qwen3-8B-Maxima_A1")
model = AutoModelForCausalLM.from_pretrained("Rumiii/Qwen3-8B-Maxima_A1", device_map="auto")
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 Settings

vLLM

How to use Rumiii/Qwen3-8B-Maxima_A1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Rumiii/Qwen3-8B-Maxima_A1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Rumiii/Qwen3-8B-Maxima_A1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Rumiii/Qwen3-8B-Maxima_A1

SGLang

How to use Rumiii/Qwen3-8B-Maxima_A1 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 "Rumiii/Qwen3-8B-Maxima_A1" \
    --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": "Rumiii/Qwen3-8B-Maxima_A1",
		"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 "Rumiii/Qwen3-8B-Maxima_A1" \
        --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": "Rumiii/Qwen3-8B-Maxima_A1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use Rumiii/Qwen3-8B-Maxima_A1 with Docker Model Runner:
```
docker model run hf.co/Rumiii/Qwen3-8B-Maxima_A1
```

Qwen3-8B fine-tuned on Toucan-1.5M (tool calling)

Merged 16-bit QLoRA fine-tune of Qwen3-8B on a stratified 1,000-sample slice of the SFT config of Agent-Ark/Toucan-1.5M.

Training data

Stratified across Toucan's pipeline stages so the model learns both to call tools and to decline when none fit:

Stage	Rows	Purpose
multi-turn	400	Extended interactions, densest tool use
single-turn-original	250	Basic single-shot tool calling
irrelevant	200	Declining when available tools don't match (zero tool calls by design)
single-turn-diversify	150	Reworded task variants

Median 2 tool calls per row (max 19), median 1,875 tokens. Examples over 4,096 tokens were dropped rather than truncated, to avoid teaching the model to stop mid-trajectory.

Training setup

QLoRA, 4-bit, rank 16, alpha 16, all attention + MLP projections
1 epoch, 125 steps, effective batch 8, lr 2e-4 cosine
Loss on assistant turns only (tool outputs masked)
Single Tesla T4, fp16, ~1h47m
Final training loss 1.18 (from 2.17)

Verified behaviour

Two-sided probe on held-out examples:

Calls when it should — given PubChem tools and asked for caffeine properties, emitted a valid <tool_call> with the correct tool name and arguments.
Declines when it should — given only Figma tools for an academic citation query, emitted no tool call, named the mismatch, and redirected to appropriate resources.

Caveats

Small-scale run: 1,000 samples against a 119k-row config drawn from a 1.5M-row dataset. Not benchmarked against base Qwen3-8B on BFCL V3 or MCP-Universe, so any capability claim is unverified. This validates the training recipe rather than demonstrating a capability gain over the base model.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Rumiii/qwen3-8b-toucan"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

tools = [{"type": "function", "function": {
    "name": "get_weather",
    "description": "Get current weather for a city",
    "parameters": {"type": "object",
                   "properties": {"city": {"type": "string"}},
                   "required": ["city"]}}}]

msgs = [{"role": "user", "content": "What's the weather in Hyderabad?"}]
text = tok.apply_chat_template(
    msgs, tools=tools, tokenize=False, add_generation_prompt=True
)
out = model.generate(
    **tok(text, return_tensors="pt").to(model.device),
    max_new_tokens=512, temperature=0.6, top_p=0.95, top_k=20
)
print(tok.decode(out[0], skip_special_tokens=False))

Sampling: temperature 0.6, top_p 0.95, top_k 20. Avoid greedy decoding (Qwen3 guidance — it can cause repetition loops).

Downloads last month: 114

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for Rumiii/Qwen3-8B-Maxima_A1

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Finetuned

(2093)

this model

Rumiii
/

Qwen3-8B-Maxima_A1