Qwen3-4B-Instruct-2507 Distilled from Kimi K2

This is a repository which hosts the model files of the Model Distillation performed from the parent model Kimi K2 to the student model Qwen3-4B-Instruct-2507, with the aim of further advancing the student's agentic capabilities, particularly multi-step tool calling.

Model Overview

Training Data

Toucan-1.5M contains over 1.5 million agentic trajectories synthesized from 495 real-world Model Context Protocol (MCP) servers, spanning more than 2,000 tools. It includes single-turn and multi-turn interactions, as well as sequential and parallel tool calls with real tool execution.

Training used the Kimi-K2 configuration of the dataset. A total of 9,168 samples were used for a single epoch, taken from all four subsets in the following distribution:

How to use it

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Rumiii/Qwimi-4B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="cuda"
)

messages = [
    {"role": "user", "content": "What is the capital of France?"}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    do_sample=False
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Only a portion of the full dataset was used. Samples longer than 4,096 tokens were excluded (no sample was truncated), as were multi-turn samples that contain tool calls but no declared tools. The trajectories follow the standard Qwen3 tool-calling chat template.

Note

This distillation was performed through an existing dataset. The Kimi K2 responses were not generated live for this model; the trajectories come from Agent-Ark/Toucan-1.5M, which was already available on Hugging Face.