modrill/math-nothink-q4b-20260908
The modrill/math-nothink-q4b-20260908 is a 4 billion parameter model, based on Qwen3-4B-Base, specifically developed by modrill as a source expert for mathematical reasoning tasks. This model is not a general-purpose chatbot but is optimized for scoring on complex mathematical problems, demonstrating a significant performance improvement over its base model on the AIME24+25 Exact-240 benchmark. It features a context length of 32768 tokens and is designed for task-vector transfer in advanced mathematical applications.
Loading preview...
Overview
The modrill/math-nothink-q4b-20260908 is a specialized 4 billion parameter model, developed by modrill, serving as a "source expert" for advanced mathematical reasoning. It is derived from Qwen/Qwen3-4B-Base and has been fine-tuned using a LoRA adapter. This model is explicitly not a general-purpose chatbot, but rather an endpoint designed for scoring on specific mathematical challenges.
Key Capabilities & Performance
This model excels in mathematical problem-solving, particularly as evaluated on the AIME24+25 Exact-240 benchmark. It achieved a score of 49/240, which is a substantial improvement over its Same-run Base model's score of 22/240. The training involved 165 updates over 10.7 million tokens, utilizing a LoRA configuration with r64/α128 and a TPU 65536 setup.
Intended Use Case
This model is intended for research and development in mathematical task-vector transfer, specifically for the ICLR 2027 task. Its primary function is to provide a score for mathematical problems, making it suitable for applications requiring precise mathematical evaluation rather than conversational interaction. Developers can load the model using AutoModelForCausalLM and AutoTokenizer from the Hugging Face transformers library.