Open R1

DeepScaleR is a 1.5-billion-parameter language model fine-tuned from DeepSeek-R1-Distilled-Qwen-1.5B using distributed reinforcement learning and a novel iterative context-lengthening strategy that gradually increases its context window from 8K to 24K tokens during training. It was trained on ~40,000 carefully curated mathematical problems drawn from competition-level datasets like AIME (1984–2023), AMC (pre-2023), Omni-MATH, and STILL. DeepScaleR achieves 43.1% accuracy on AIME 2024, a roughly 14.3 percentage point boost over the base model, and surpasses the performance of the proprietary O1-Preview model despite its much smaller size. It also posts strong results on a suite of math benchmarks (e.g., MATH-500, AMC 2023, Minerva Math, OlympiadBench), demonstrating that small, efficient models tuned with RL can match or exceed larger baselines on reasoning tasks.

Phi-4-mini-reasoning is a 3.8-billion parameter transformer-based language model optimized for mathematical reasoning and step-by-step problem solving in environments with constrained computing or latency. Fine-tuned with synthetic data generated by the DeepSeek-R1 model, it balances efficiency with advanced reasoning ability. Trained on over one million diverse math problems spanning multiple levels of difficulty from middle school to Ph.D. level, Phi-4-mini-reasoning outperforms its base model on long sentence generation across various evaluations and surpasses larger models like OpenThinker-7B, Llama-3.2-3B-instruct, and DeepSeek-R1. It features a 128K-token context window and supports function calling, enabling integration with external tools and APIs. Phi-4-mini-reasoning can be quantized using Microsoft Olive or Apple MLX Framework for deployment on edge devices such as IoT, laptops, and mobile devices.