What is tensorhydra/qwen3-8b-aimo3-tir?
This is an 8 billion parameter Qwen-8B model that has been fine-tuned using LoRA for tool-integrated reasoning. The LoRA adapters are already merged into the base model and saved in SafeTensors format, meaning you don't need PEFT for deployment.
Key Capabilities & Features
- Tool-Integrated Reasoning: Specialized in solving problems that require external tools or step-by-step logical deduction.
- Mathematical Problem Solving: Optimized for mathematical reasoning and complex problem-solving, trained on the AIMO3 competition dataset.
- Efficient Deployment: Provided as a merged model in SafeTensors (BF16) format, simplifying loading and inference.
- Flexible Quantization: Supports 8-bit and 4-bit quantization for reduced VRAM usage (down to ~4GB).
- Context Length: Features a maximum context length of 8192 tokens.
Training Details
The model was fine-tuned on a single NVIDIA H100 GPU using BFloat16 precision for 2 epochs, with a max sequence length of 8192. The training data consists of problem-solution pairs from the AIMO3 Tool-Integrated Reasoning Dataset, which was synthesized by GPT-OSS-120B.
Should you use this model?
This model is ideal for use cases requiring strong mathematical reasoning, problem-solving, and tool-integrated step-by-step reasoning. It's particularly well-suited for educational and research applications in these domains. Its merged LoRA configuration makes it convenient for production deployment where PEFT dependencies are undesirable. However, it may not generalize well to tasks outside its specialized reasoning domain due to its narrow fine-tuning.