Qwen3-8B-Maxima_A1

Qwen3-8B fine-tuned for tool calling and agentic workflows on a 5,000-sample subset of the SFT split of Toucan-1.5M.

Given a set of tool definitions, the model emits <tool_call> JSON, chains calls across steps using tool results, and answers directly when no tool is relevant. It is a small-scale proof of concept, released as merged 16-bit weights.

Training

Usage

The 16-bit weights are about 15 GB. On smaller GPUs, load the model in 4-bit (for example with bitsandbytes).

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Rumiii/Qwen3-8B-Maxima_A1"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")

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

messages = [{"role": "user", "content": "What's the weather in Hyderabad?"}]
prompt = tok.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=False).split("<|im_end|>")[0])

Expected output:

<tool_call>
{"name": "get_weather", "arguments": {"city": "Hyderabad"}}
</tool_call>

To continue after running the tool, append the call and its result to messages, rebuild the prompt with apply_chat_template, and generate again:

messages += [
    {"role": "assistant", "content": "", "tool_calls": [
        {"type": "function", "function": {"name": "get_weather", "arguments": {"city": "Hyderabad"}}}]},
    {"role": "tool", "content": '{"temp_c": 31, "condition": "Partly cloudy"}'},
]

Limitations

• Trained on 5,000 samples only and not evaluated on public benchmarks.
• The training data is synthetic, generated from community MCP servers between June and September 2025.
• Always validate tool names and arguments before executing a tool call.

Acknowledgements

Data: Toucan-1.5M (Apache 2.0). Base model: Qwen3-8B (Apache 2.0).