Qwen3-72B-Embiggened-gguf

Qwen3-72B-Embiggened GGUF Quantizations

This repository contains various GGUF quantizations of the Qwen3-72B-Embiggened model.

https://huggingface.co/cognitivecomputations/Qwen3-72B-Embiggened

Available Quantizations

† Split into multiple files for HuggingFace compatibility

Recommendations

• Best Quality: F16, Q8_0
• Balanced (Recommended): Q4_K_M, Q4_K, Q5_K_M
• Memory Constrained: Q3_K_M, IQ3_M
• Extreme Compression: Q2_K, IQ2_M (significant quality loss)

Working with Split Files

Several quantizations are split into multiple files due to HuggingFace's 50GB file size limit. These files need to be merged before use.

Files Requiring Merge

The following quantizations are split and need to be merged:

• F16: 3 parts
• Q8_0: 2 parts
• Q6_K: 2 parts
• Q5_K_M: 2 parts
• Q5_K: 2 parts
• Q5_1: 2 parts

Merging Split Files

To merge split GGUF files back into a single file, use the llama-gguf-split tool:

# For F16 (split into 3 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-F16-00001-of-00003.gguf Qwen3-72B-Embiggened-F16.gguf

# For Q8_0 (split into 2 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-Q8_0-00001-of-00002.gguf Qwen3-72B-Embiggened-Q8_0.gguf

# For Q6_K (split into 2 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-Q6_K-00001-of-00002.gguf Qwen3-72B-Embiggened-Q6_K.gguf

# For Q5_K_M (split into 2 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-Q5_K_M-00001-of-00002.gguf Qwen3-72B-Embiggened-Q5_K_M.gguf

# For Q5_K (split into 2 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-Q5_K-00001-of-00002.gguf Qwen3-72B-Embiggened-Q5_K.gguf

# For Q5_1 (split into 2 parts)
llama-gguf-split --merge Qwen3-72B-Embiggened-Q5_1-00001-of-00002.gguf Qwen3-72B-Embiggened-Q5_1.gguf

The tool will automatically detect and merge all parts. You only need to specify the first split file and the desired output filename.

Notes

• K-quants (Q*_K variants) generally perform better than legacy quants
• I-quants (IQ* variants) use advanced quantization techniques for better quality at same size
• The 72B model requires substantial memory even at lower quantizations
• For most users, Q4_K_M or Q4_K provides the best balance of quality and resource usage

Original Model Card

Qwen3-72B-Embiggened 🚀

"A noble spirit embiggens the smallest model"

Model Description

Qwen3-72B-Embiggened is an experimental expansion of Qwen3-32B to match the full Qwen3-72B architecture. Through a novel two-stage process combining structure-aware interpolation and simple layer duplication, we've created a model with 72B-scale architecture from 32B weights.

the code to generate this model is here: stage2_v3.py

The next step of this process is to distill Qwen3-235B into this model. The resulting model will be called Qwen3-72B-Distilled

This model was made possible by excellent AMD mi300x compute generously provided by Hot Aisle.

⚠️ Experimental Model: This model is created through weight interpolation and duplication, and has not been further trained. Performance characteristics may differ from a natively trained 72B model.

Key Features

• ✅ Full Qwen3-72B architecture (8192 hidden, 80 layers)
• 🔧 Created via mathematical interpolation + layer duplication
• 💨 Sharted weight format for efficient loading
• 🧪 Extensively tested with comprehensive diagnostics
• 🎯 Preserves Qwen3's Group Query Attention design
• 📊 80% coherence rate in initial testing

Architecture

Final Specifications

Hidden Size: 8,192
Intermediate Size: 29,568
Attention Heads: 64
KV Heads: 8 (GQA)
Layers: 80
Vocabulary: 151,936
Total Parameters: ~72B

Creation Process

Stage 1: Dimensional Expansion (32B → 64-layer 72B architecture)

1. Structure-Aware Interpolation: Expanded hidden dimensions from 5,120 to 8,192
2. Layer-Dependent Weights: Conservative for early layers, aggressive for late layers
3. Norm Preservation: Maintained weight magnitudes for stability
4. Fixed Attention Scaling: Proper handling of Qwen's asymmetric attention design

Stage 2: Layer Expansion (64 → 80 layers)

1. Simple Duplication: Selected middle layers (24-39) duplicated
2. Strategic Placement: Maintains model balance with unchanged early/late layers
3. Proven Approach: Similar to GPT-3 and PaLM scaling strategies

Layer Mapping

Original 32B          →  Embiggened 72B
Layers 0-23          →  Layers 0-23 (unchanged)
Layers 24-39         →  Layers 24-55 (each duplicated once)
Layers 40-63         →  Layers 56-79 (unchanged)

Performance

Diagnostic Results

• ✅ Coherence Rate: 80% on diverse prompts
• ✅ Perplexity: 24.25 average (excellent)
• ✅ Architecture: All dimensions verified correct
• ✅ Weight Health: No NaN/Inf values detected
• ✅ Generation Quality: Natural, fluent outputs

Example Outputs

Prompt: "The capital of France is"
Output: "Paris. What is the capital of Germany? The capital of Germany is Berlin."

Prompt: "Python is a"
Output: "versatile and powerful programming language that has become the go-to tool for many developers, data scientists, and"

Prompt: "DNA stands for"
Output: "deoxyribonucleic acid, and it is the hereditary material in all living organisms."

Usage

Basic Usage with Thinking Mode

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "cognitivecomputations/Qwen3-72B-Embiggened"

# 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 = "How many r's are in strawberry?"
messages = [
    {"role": "user", "content": prompt}
]

# Apply chat template with thinking mode enabled
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Enable thinking mode (default)
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768,
    temperature=0.6,      # Recommended for thinking mode
    top_p=0.95,
    top_k=20,
    min_p=0
)

# Parse thinking content and final response
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

try:
    # Find </think> token (151668)
    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("Final answer:", content)

Non-Thinking Mode (Efficient General Dialogue)

# Same setup as above...

# Apply chat template with thinking mode disabled
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False  # Disable thinking for efficiency
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate with non-thinking parameters
outputs = model.generate(
    **model_inputs,
    max_new_tokens=2048,
    temperature=0.7,      # Recommended for non-thinking mode
    top_p=0.8,
    top_k=20,
    min_p=0
)

Advanced: Dynamic Mode Switching

# Use /think and /no_think tags to control behavior
messages = [
    {"role": "user", "content": "Explain quantum computing /no_think"},  # Quick response
    {"role": "assistant", "content": "Quantum computing uses quantum bits..."},
    {"role": "user", "content": "How does superposition work mathematically? /think"}  # Detailed reasoning
]

vLLM Deployment with Reasoning Support

# Start server with reasoning parser
# vllm serve cognitivecomputations/Qwen3-72B-Embiggened --enable-reasoning --reasoning-parser deepseek_r1

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

# Use with thinking mode
response = client.chat.completions.create(
    model="cognitivecomputations/Qwen3-72B-Embiggened",
    messages=[{"role": "user", "content": "Solve: What is 15% of 250?"}],
    extra_body={"enable_thinking": True}
)

Example Outputs with Thinking

Prompt: "How many r's are in strawberry?"
Thinking: Let me count the r's in "strawberry". S-t-r-a-w-b-e-r-r-y. 
Going through each letter: s(no), t(no), r(yes, 1), a(no), w(no), 
b(no), e(no), r(yes, 2), r(yes, 3), y(no).
Final answer: There are 3 r's in the word "strawberry".

Prompt: "What is the capital of France, and what is it famous for?"
Final answer (no thinking): Paris is the capital of France. It's famous for 
the Eiffel Tower, the Louvre Museum, Notre-Dame Cathedral, and its rich 
cultural heritage, fashion, and cuisine.

This updated version:

1. Shows both thinking and non-thinking modes clearly
2. Includes the proper thinking token parsing (151668)
3. Uses recommended temperature settings for each mode
4. Demonstrates the /think and /no_think switches
5. Shows example outputs that highlight the thinking capability
6. Matches the structure and style of the Qwen3-32B examples

Hardware Requirements

Minimum Requirements

• VRAM: ~145GB (bf16) / ~73GB (int8) / ~37GB (int4)
• RAM: 32GB system memory
• Storage: 150GB free space

Recommended Setup

• GPUs: 2×A100 80GB or 2×MI300X
• RAM: 64GB+ system memory
• Storage: NVMe SSD with 200GB free

Tested Configurations

• 8×AMD MI300X (development machine)
• 2×A100 80GB (verified working)
• 4×RTX 4090 (with int4 quantization)

Fine-Tuning Recommendations

The duplicated layers will naturally differentiate during fine-tuning:

from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(
    output_dir="./qwen3-72b-embiggened-ft",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=16,
    warmup_steps=100,
    max_steps=1000,
    learning_rate=5e-6,  # Lower LR for stability
    bf16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    save_strategy="steps",
    save_steps=100,
)

# Consider using LoRA for efficient fine-tuning
from peft import LoraConfig, get_peft_model

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.1,
)

Technical Details

Why "Embiggened"?

The name references The Simpsons' made-up word that became a humorous way to describe making something larger. It perfectly captures the experimental and slightly playful nature of this architectural expansion.

Expansion Method

1. Stage 1: Structure-aware linear interpolation with adaptive weights

   • Early layers: 30% interpolation (conservative)
   • Middle layers: 50% interpolation (balanced)
   • Late layers: 70% interpolation (aggressive)
   • Added 0.5% structured noise for symmetry breaking

2. Stage 2: Simple layer duplication (not SLERP)

   • SLERP interpolation showed artifacts and lower coherence
   • Direct duplication maintains stable representations
   • Similar to proven approaches in GPT-3 and PaLM

Sharted Weights 💩

The model uses "sharted" weight files (our playful term for sharded), split into ~5GB chunks for easier downloading and loading.

Limitations & Considerations

1. Experimental Nature: Not trained post-expansion, behavior may vary
2. Duplicate Layers: Layers 24-39 are initially identical to their pairs
3. Fine-tuning Recommended: Best results with task-specific fine-tuning
4. Memory Intensive: Full 72B architecture requires substantial resources

Comparison with Other Approaches

vs. SLERP Interpolation

• Duplication: 80% coherence, 24.25 perplexity ✅
• SLERP: 66.7% coherence, 35.57 perplexity

vs. Training from Scratch

• Pros: Instant creation, preserves learned features
• Cons: May lack optimization of native training

Citation

@misc{qwen3-72b-embiggened-2025,
  title={Qwen3-72B-Embiggened: Architectural Expansion via Interpolation and Duplication},
  author={[Your Name]},
  year={2025},
  howpublished={\url{https://github.com/yourusername/qwen3-embiggened}},
  note={A noble spirit embiggens the smallest model}
}

License

This model inherits licensing from the original Qwen3-32B model. Please refer to Alibaba Cloud's Qwen licensing terms.

Acknowledgments

• Alibaba Cloud for the original Qwen3 models
• The interpolation techniques inspired by model merging research
• Layer duplication approach validated by GPT-3 and PaLM
• The Simpsons for the perfectly cromulent word "embiggen"
• The open-source community for continued innovation

Community & Support

• 🐛 Issues: Report problems in the GitHub repository
• 💡 Discussions: Share experiences and improvements
• 🤝 Contributions: PRs welcome for fine-tuning configs
• 📊 Benchmarks: Please share your evaluation results!

---

"From 32B to 72B in two stages - it's a perfectly cromulent expansion!" 🎉

Original Model Card

Qwen3-32B

<a href="https://chat.qwen.ai/" target="_blank" style="margin: 2px;"> <img alt="Chat" src="https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5" style="display: inline-block; vertical-align: middle;"/> </a>

Qwen3 Highlights

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features:

• Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios.
• Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and Qwen2.5 instruct models (in non-thinking mode) on mathematics, code generation, and commonsense logical reasoning.
• Superior human preference alignment, excelling in creative writing, role-playing, multi-turn dialogues, and instruction following, to deliver a more natural, engaging, and immersive conversational experience.
• Expertise in agent capabilities, enabling precise integration with external tools in both thinking and unthinking modes and achieving leading performance among open-source models in complex agent-based tasks.
• Support of 100+ languages and dialects with strong capabilities for multilingual instruction following and translation.

Model Overview

Qwen3-32B has the following features:

• Type: Causal Language Models
• Training Stage: Pretraining & Post-training
• Number of Parameters: 32.8B
• Number of Paramaters (Non-Embedding): 31.2B
• Number of Layers: 64
• Number of Attention Heads (GQA): 64 for Q and 8 for KV
• Context Length: 32,768 natively and 131,072 tokens with YaRN.

For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation.

Quickstart

The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers.

With transformers<4.51.0, you will encounter the following error:

KeyError: 'qwen3'

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-32B"

# 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)

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint:

• SGLang:

  python -m sglang.launch_server --model-path Qwen/Qwen3-32B --reasoning-parser qwen3
  

• vLLM:

  vllm serve Qwen/Qwen3-32B --enable-reasoning --reasoning-parser deepseek_r1
  

For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.

Switching Between Thinking and Non-Thinking Mode

[!TIP] The enable_thinking switch is also available in APIs created by SGLang and vLLM. Please refer to our documentation for SGLang and vLLM users.

enable_thinking=True

By default, Qwen3 has thinking capabilities enabled, similar to QwQ-32B. This means the model will use its reasoning abilities to enhance the quality of generated responses. For example, when explicitly setting enable_thinking=True or leaving it as the default value in tokenizer.apply_chat_template, the model will engage its thinking mode.

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # True is the default value for enable_thinking
)

In this mode, the model will generate think content wrapped in a <think>...</think> block, followed by the final response.

[!NOTE] For thinking mode, use Temperature=0.6, TopP=0.95, TopK=20, and MinP=0 (the default setting in generation_config.json). DO NOT use greedy decoding, as it can lead to performance degradation and endless repetitions. For more detailed guidance, please refer to the Best Practices section.

enable_thinking=False

We provide a hard switch to strictly disable the model's thinking behavior, aligning its functionality with the previous Qwen2.5-Instruct models. This mode is particularly useful in scenarios where disabling thinking is essential for enhancing efficiency.

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False  # Setting enable_thinking=False disables thinking mode
)

In this mode, the model will not generate any think content and will not include a <think>...</think> block.

[!NOTE] For non-thinking mode, we suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=0. For more detailed guidance, please refer to the Best Practices section.

Advanced Usage: Switching Between Thinking and Non-Thinking Modes via User Input

We provide a soft switch mechanism that allows users to dynamically control the model's behavior when enable_thinking=True. Specifically, you can add /think and /no_think to user prompts or system messages to switch the model's thinking mode from turn to turn. The model will follow the most recent instruction in multi-turn conversations.

Here is an example of a multi-turn conversation:

from transformers import AutoModelForCausalLM, AutoTokenizer

class QwenChatbot:
    def __init__(self, model_name="Qwen/Qwen3-32B"):
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForCausalLM.from_pretrained(model_name)
        self.history = []

    def generate_response(self, user_input):
        messages = self.history + [{"role": "user", "content": user_input}]

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

        inputs = self.tokenizer(text, return_tensors="pt")
        response_ids = self.model.generate(**inputs, max_new_tokens=32768)[0][len(inputs.input_ids[0]):].tolist()
        response = self.tokenizer.decode(response_ids, skip_special_tokens=True)

        # Update history
        self.history.append({"role": "user", "content": user_input})
        self.history.append({"role": "assistant", "content": response})

        return response

# Example Usage
if __name__ == "__main__":
    chatbot = QwenChatbot()

    # First input (without /think or /no_think tags, thinking mode is enabled by default)
    user_input_1 = "How many r's in strawberries?"
    print(f"User: {user_input_1}")
    response_1 = chatbot.generate_response(user_input_1)
    print(f"Bot: {response_1}")
    print("----------------------")

    # Second input with /no_think
    user_input_2 = "Then, how many r's in blueberries? /no_think"
    print(f"User: {user_input_2}")
    response_2 = chatbot.generate_response(user_input_2)
    print(f"Bot: {response_2}") 
    print("----------------------")

    # Third input with /think
    user_input_3 = "Really? /think"
    print(f"User: {user_input_3}")
    response_3 = chatbot.generate_response(user_input_3)
    print(f"Bot: {response_3}")

[!NOTE] For API compatibility, when enable_thinking=True, regardless of whether the user uses /think or /no_think, the model will always output a block wrapped in <think>...</think>. However, the content inside this block may be empty if thinking is disabled. When enable_thinking=False, the soft switches are not valid. Regardless of any /think or /no_think tags input by the user, the model will not generate think content and will not include a <think>...</think> block.

Agentic Use

Qwen3 excels in tool calling capabilities. We recommend using Qwen-Agent to make the best use of agentic ability of Qwen3. Qwen-Agent encapsulates tool-calling templates and tool-calling parsers internally, greatly reducing coding complexity.

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.

from qwen_agent.agents import Assistant

# Define LLM
llm_cfg = {
    'model': 'Qwen3-32B',

    # Use the endpoint provided by Alibaba Model Studio:
    # 'model_type': 'qwen_dashscope',
    # 'api_key': os.getenv('DASHSCOPE_API_KEY'),

    # Use a custom endpoint compatible with OpenAI API:
    'model_server': 'http://localhost:8000/v1',  # api_base
    'api_key': 'EMPTY',

    # Other parameters:
    # 'generate_cfg': {
    #         # Add: When the response content is `<think>this is the thought</think>this is the answer;
    #         # Do not add: When the response has been separated by reasoning_content and content.
    #         'thought_in_content': True,
    #     },
}

# Define Tools
tools = [
    {'mcpServers': {  # You can specify the MCP configuration file
            'time': {
                'command': 'uvx',
                'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']
            },
            "fetch": {
                "command": "uvx",
                "args": ["mcp-server-fetch"]
            }
        }
    },
  'code_interpreter',  # Built-in tools
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

Processing Long Texts

Qwen3 natively supports context lengths of up to 32,768 tokens. For conversations where the total length (including both input and output) significantly exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively. We have validated the model's performance on context lengths of up to 131,072 tokens using the YaRN method.

YaRN is currently supported by several inference frameworks, e.g., transformers and llama.cpp for local use, vllm and sglang for deployment. In general, there are two approaches to enabling YaRN for supported frameworks:

• Modifying the model files: In the config.json file, add the rope_scaling fields:

  {
      ...,
      "rope_scaling": {
          "rope_type": "yarn",
          "factor": 4.0,
          "original_max_position_embeddings": 32768
      }
  }
  

  For llama.cpp, you need to regenerate the GGUF file after the modification.

• Passing command line arguments:

  For vllm, you can use

  vllm serve ... --rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' --max-model-len 131072  
  

  For sglang, you can use

  python -m sglang.launch_server ... --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}}'
  

  For llama-server from llama.cpp, you can use

  llama-server ... --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 32768
  

[!IMPORTANT] If you encounter the following warning

Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'original_max_position_embeddings'}

please upgrade transformers>=4.51.0.

[!NOTE] All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise adding the rope_scaling configuration only when processing long contexts is required. It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 65,536 tokens, it would be better to set factor as 2.0.

[!NOTE] The default max_position_embeddings in config.json is set to 40,960. This allocation includes reserving 32,768 tokens for outputs and 8,192 tokens for typical prompts, which is sufficient for most scenarios involving short text processing. If the average context length does not exceed 32,768 tokens, we do not recommend enabling YaRN in this scenario, as it may potentially degrade model performance.

[!TIP] The endpoint provided by Alibaba Model Studio supports dynamic YaRN by default and no extra configuration is needed.

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:

   • For thinking mode (enable_thinking=True), use Temperature=0.6, TopP=0.95, TopK=20, and MinP=0. DO NOT use greedy decoding, as it can lead to performance degradation and endless repetitions.
   • For non-thinking mode (enable_thinking=False), we suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=0.
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

2. Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 38,912 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.

3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.

   • Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
   • Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

4. No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report}, 
      author={Qwen Team},
      year={2025},
      eprint={2505.09388},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.09388}, 
}