cyankiwi / IQuest-Coder-V1-40B-Loop-Instruct-AWQ-4bit

📘 Blog •  📄 Technical Report

IQuest-Coder-V1 Model Family

Clarification: Regarding the Performance of IQuest-Coder-V1

Sampling Parameters

For the IQuest-Coder-V1-Instruct: We suggest using Temperature=0.6, TopP=0.85, TopK=20.

IQuest-Coder-V1 Highlights

IQuest-Coder-V1 is a new family of code large language models (LLMs) designed to advance autonomous software engineering and code intelligence. Built on the innovative code-flow multi-stage training paradigm, IQuest-Coder-V1 captures the dynamic evolution of software logic, delivering state-of-the-art performance across critical dimensions:

• State-of-the-Art Performance : Achieves leading results on SWE-Bench Verified (76.2%), BigCodeBench (49.9%), LiveCodeBench v6 (81.1%), and other major coding benchmarks, surpassing competitive models across agentic software engineering, competitive programming, and complex tool use.
• Code-Flow Training Paradigm : Moving beyond static code representations, our models learn from repository evolution patterns, commit transitions, and dynamic code transformations to understand real-world software development processes.
• Dual Specialization Paths : Bifurcated post-training delivers two specialized variants—Thinking models (utilizing reasoning-driven RL for complex problem-solving) and Instruct models (optimized for general coding assistance and instruction-following).
• Efficient Architecture : The IQuest-Coder-V1-Loop variant introduces a recurrent mechanism that optimizes the trade-off between model capacity and deployment footprint.
• Native Long Context : All models natively support up to 128K tokens without requiring additional scaling techniques.

Model Overview

The IQuest-Coder-V1 series includes models ranging from 7B to 40B parameters, with both standard and Loop variants:

Architecture Features:

• Grouped Query Attention (GQA) for efficient inference
• Native 128K context length support
• Vocabulary size: 76,800 tokens
• Loop variants use recurrent transformer design with shared parameters across two iterations

For more details, please refer to our Technical Report, GitHub.

Quickstart

IQuest-Coder-V1 uses custom modeling code via Hugging Face's auto_map feature. We recommend using transformers==4.56.0.

Basic Usage with Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "IQuestLab/IQuest-Coder-V1-40B-Instruct"

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="cuda:0",
    trust_remote_code=True,
)

# Prepare the input
prompt = "Write a Python function to calculate the Fibonacci sequence using dynamic programming."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=8192
)
generated_ids = generated_ids[0][len(model_inputs.input_ids[0]):]
response = tokenizer.decode(generated_ids, skip_special_tokens=True)

print(response)

Using Thinking Models

For complex reasoning tasks, use the Thinking variant:

model_name = "IQuest/IQuest-Coder-V1-40B-Thinking"

# The Thinking model includes explicit reasoning traces
# Use similar code as above, but expect longer, more detailed responses
# with step-by-step problem decomposition

Deployment with vLLM

For production deployment, you can use vLLM to create an OpenAI-compatible API endpoint. Please refer to the vLLM PR for implementation details.

vllm serve IQuestLab/IQuest-Coder-V1-40B-Instruct --tensor-parallel-size 8

For Thinking models with reasoning support:

vllm serve IQuestLab/IQuest-Coder-V1-40B-Thinking --reasoning-parser qwen3 --tensor-parallel-size 8

Evaluation Results

Limitations

• Reasoning vs. Efficiency Trade-off : Thinking models provide superior reasoning but generate longer responses; Instruct models are more efficient for straightforward tasks.
• Code Execution : Models generate code but do not execute it; always validate outputs in sandboxed environments.
• Domain Specificity : While trained on diverse codebases, performance may vary on highly specialized or proprietary frameworks.
• Factuality : Models may generate plausible but incorrect code; verify critical implementations thoroughly.

Citation

If you find our work helpful, please cite:

@article{iquest-coder-v1-2025,
  title={IQuest-Coder-V1 Technical Report},
  author={IQuest Coder Team},
  url={https://github.com/IQuestLab/IQuest-Coder-V1/blob/main/papers/IQuest_Coder_Technical_Report.pdf}
  year={2025}
}
@article{codescaling,
  title={Scaling Laws for Code: Every Programming Language Matters},
  author={Yang, Jian and Guo, Shawn and Jing, Lin and Zhang, Wei and Liu, Aishan and Hao, Chuan and Li, Zhoujun and Zhao, Wayne Xin and Liu, Xianglong and Lv, Weifeng and others},
  journal={arXiv preprint arXiv:2512.13472},
  year={2025}
}
@article{close_the_loop,
  title={Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing},
  author={Yuwen Li, Wei Zhang, Zelong Huang, Mason Yang, Jiajun Wu, Shawn Guo, Huahao Hu, Lingyi Sun, Jian Yang, Mingjie Tang, Byran Dai},
  journal={arXiv preprint arXiv:2512.23611},
  year={2025}
}
@article{loopcoder,
  title={LoopCoder: Scaling Code Intelligence via Looped Language Models},
  author={Jian Yang, Wei Zhang, Shawn Guo, Yizhi Li, Lin Jing, Zhengmao Ye, Shark Liu, Yuyang Song, Jiajun Wu, Che Liu, T. Zheng, Siwei Wu, L. Liao, X. Ma, Chuan Hao, Ran Tao, Yan Xing, Jianzhou Wang, Mingjie Tang, Aishan Liu, Zhoujun Li, Xianglong Liu, Weifeng Lv1, Bryan Dai},
  year={2025}
}
@article{swe_compress,
  title={Context as a Tool: Context Management for Long-Horizon SWE-Agents},
  author={hukai Liu, Jian Yang, Bo Jiang, Yizhi Li, Jinyang Guo, Xianglong Liu, Bryan Dai},
  journal={arXiv preprint arXiv:2512.22087},
  year={2025}
}