Qwen3-4b-tcomanr-merge-v2

Ties merged COde MAth aNd Reasoning model

This is a merge of pre-trained language models created using mergekit.

Merge Details

This model aims to combine the code and math capabilities by merging Qwen 3 2507 with multiple Qwen 3 finetunes.

How to run

You can run this model by using multiple interface choices

transformers

As the qwen team suggested to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ertghiu256/Qwen3-4b-tcomanr-merge-v2"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

vllm

Run this command

vllm serve ertghiu256/Qwen3-4b-tcomanr-merge-v2 --enable-reasoning --reasoning-parser deepseek_r1

Sglang

Run this command

python -m sglang.launch_server --model-path ertghiu256/Qwen3-4b-tcomanr-merge-v2 --reasoning-parser deepseek-r1

llama.cpp

Run this command

llama-server --hf-repo ertghiu256/Qwen3-4b-tcomanr-merge-v2

or

llama-cli --hf ertghiu256/Qwen3-4b-tcomanr-merge-v2

ollama

Run this command

ollama run hf.co/ertghiu256/Qwen3-4b-tcomanr-merge-v2:Q8_0

lm studio

Search

ertghiu256/Qwen3-4b-tcomanr-merge-v2

in the lm studio model search list then download

Recomended parameters

temp: 0.6
num_ctx: ≥8192
top_p: 0.95
top_k: 30
Repeat Penalty: 1.2

Merge Method

This model was merged using the TIES merge method using Qwen/Qwen3-4B-Thinking-2507 as a base.

Models Merged

The following models were included in the merge:

• POLARIS-Project/Polaris-4B-Preview
• ertghiu256/qwen-3-4b-mixture-of-thought
• ertghiu256/qwen3-multi-reasoner
• ertghiu256/Qwen3-Hermes-4b
• ertghiu256/deepseek-r1-0528-distilled-qwen3
• Tesslate/UIGEN-T3-4B-Preview-MAX
• ertghiu256/qwen3-math-reasoner
• ValiantLabs/Qwen3-4B-Esper3
• huihui-ai/Huihui-Qwen3-4B-Thinking-2507-abliterated
• ertghiu256/qwen3-4b-code-reasoning
• ValiantLabs/Qwen3-4B-ShiningValiant3

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: ertghiu256/qwen3-math-reasoner
    parameters:
      weight: 0.8
  - model: ertghiu256/qwen3-4b-code-reasoning
    parameters:
      weight: 0.8
  - model: ertghiu256/qwen-3-4b-mixture-of-thought
    parameters:
      weight: 0.9
  - model: POLARIS-Project/Polaris-4B-Preview
    parameters:
      weight: 0.7
  - model: ertghiu256/qwen3-multi-reasoner
    parameters:
      weight: 0.8
  - model: ertghiu256/Qwen3-Hermes-4b
    parameters:
      weight: 0.4
  - model: ValiantLabs/Qwen3-4B-Esper3
    parameters:
      weight: 0.8
  - model: Tesslate/UIGEN-T3-4B-Preview-MAX
    parameters:
      weight: 0.8
  - model: ValiantLabs/Qwen3-4B-ShiningValiant3
    parameters:
      weight: 0.7
  - model: ertghiu256/deepseek-r1-0528-distilled-qwen3
    parameters:
      weight: 0.2
  - model: huihui-ai/Huihui-Qwen3-4B-Thinking-2507-abliterated
    parameters:
      weight: 0.6
merge_method: ties
base_model: Qwen/Qwen3-4B-Thinking-2507
parameters:
  normalize: true
  int8_mask: true
  lambda: 1.0
dtype: float16