MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes

Changsheng Zhao, Ernie Chang, Zechun Liu, Chia-Jung Chang, Wei Wen, Chen Lai, Sheng Cao, Yuandong Tian, Raghuraman Krishnamoorthi, Yangyang Shi, Vikas Chandra

Abstract

The paradigm shift in large language models (LLMs) from instinctive responses to chain-of-thought (CoT) reasoning has fueled two prevailing assumptions: (1) reasoning capabilities only emerge in sufficiently large models, and (2) such capabilities require training on massive datasets. While the first assumption has already been challenged by recent sub-billion-parameter reasoning models such as Qwen3-0.6B and DeepSeek distilled variants, the second remains largely unquestioned. In this work, we revisit the necessity of scaling to extremely large corpora (>10T tokens) for reasoning emergence. By carefully curating and resampling open-source datasets that we identify as beneficial under our designed metrics, we demonstrate that strong reasoning abilities can emerge with far less data. Specifically, we show that only ~2T tokens of high-quality data are sufficient, and pre-training with 4.2T tokens on the dataset resampled from these ~2T tokens, followed by a established post-training procedure, enables the development of MobileLLM-R1, a series of sub-billion-parameter reasoning models that substantially outperform prior models trained on fully open-sourced data. For example, MobileLLM-R1-950M achieves an AIME score of 15.5, compared to just 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B. Remarkably, despite being trained on only 11.7% of the tokens compared to Qwen3's proprietary 36T-token corpus for pretraining, MobileLLM-R1-950M matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. To facilitate further research in this direction, we have released the complete training recipe, data sources, data mixing ratio, and model checkpoints, together with the key insights obtained throughout this study.

Introduction

In the current landscape of Large Language Models (LLMs), it is often assumed that complex reasoning is a property reserved for massive models trained on tens of trillions of tokens. While small, "distilled" models have begun to challenge the size requirement, the data requirement has remained a significant barrier. Most high-performing small models rely on proprietary, massive datasets. MobileLLM-R1 addresses this "data efficiency" gap, investigating whether top-tier reasoning can emerge in models under 1 billion parameters using only open-source data and a fraction of the typical token budget.

Methodology

The researchers implemented a multi-stage training pipeline that prioritizes data curation and high-signal samples over raw scale. The process is divided into three primary phases:

  1. Capability-Aware Pre-training: Using "Leave-One-Out" (LOO) analysis, the team identified which datasets contribute most to reasoning. They found that general web data (like FineWeb-Edu) acts as the "glue" connecting specialized domains.
  2. Mid-training Knowledge Compression: During a 100B token mid-training phase, the model uses a self-evolving strategy to identify and remove samples with negative or zero influence, focusing the training on high-quality data.
  3. Staged Post-training: A two-step alignment process was used, starting with general instruction-following (Tulu-3) followed by specialized Supervised Fine-Tuning (SFT) for math and code reasoning.
The multi-stage training pipeline for MobileLLM-R1, progressing from pre-training to specialized post-training.

Key Results

MobileLLM-R1-950M sets a new benchmark for sub-billion parameter models. It achieved an AIME score of 15.5, significantly outperforming other open-source models like OLMo-2-1.48B (0.6) and SmolLM-2-1.7B (0.3).

In terms of efficiency, the 950M model matches or exceeds the performance of Qwen3-0.6B across multiple benchmarks despite being trained on only 11.7% of the tokens. Furthermore, the model shows exceptional coding prowess, reaching 46.3% on HumanEval, the highest score among all sub-1B models tested. For on-device applications, the 140M variant sustains over 100 tokens/second on a Samsung Galaxy S22, proving the viability of real-time, local reasoning.

MobileLLM-R1 performance on HumanEval compared to other sub-billion models, showing superior accuracy relative to pre-training FLOPs.

Insights and Discussion

Several critical insights emerged from this study:

  • Quality Over Quantity: Reasoning can emerge with ~2T tokens of curated data, challenging the 10T+ token industry standard.
  • The "Glue" Effect: General web data is essential for providing the linguistic framework that connects specialized knowledge in math and code.
  • Math-to-Code Transfer: Training on mathematical data was found to improve coding logic, as evidenced by a dramatic decrease in perplexity on coding benchmarks after math-focused training.
  • RL Limitations: Reinforcement Learning (GRPO) showed diminishing returns for small models that have already been optimized with high-quality distilled SFT data, suggesting small models may lack the "exploration capacity" of their larger counterparts.

Conclusion

MobileLLM-R1 demonstrates that the limits of sub-billion parameter models have not yet been reached. By focusing on a data-centric approach and open training recipes, it is possible to create highly efficient reasoners that rival models trained on significantly larger proprietary datasets. Future work will continue to explore the co-evolution of data and models to further push the boundaries of on-device intelligence.

Summary updated

Abstract

The paradigm shift in large language models (LLMs) from instinctive responses to chain-of-thought (CoT) reasoning has fueled two prevailing assumptions: (1) reasoning capabilities only emerge in sufficiently large models, and (2) such capabilities require training on massive datasets. While the first assumption has already been challenged by recent sub-billion-parameter reasoning models such as Qwen3-0.6B and DeepSeek distilled variants, the second remains largely unquestioned. In this work, we revisit the necessity of scaling to extremely large corpora (>10T tokens) for reasoning emergence. By carefully curating and resampling open-source datasets that we identify as beneficial under our designed metrics, we demonstrate that strong reasoning abilities can emerge with far less data. Specifically, we show that only ~2T tokens of high-quality data are sufficient, and pre-training with 4.2T tokens on the dataset resampled from these ~2T tokens, followed by a established post-training procedure, enables the development of MobileLLM-R1, a series of sub-billion-parameter reasoning models that substantially outperform prior models trained on fully open-sourced data. For example, MobileLLM-R1-950M achieves an AIME score of 15.5, compared to just 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B. Remarkably, despite being trained on only 11.7% of the tokens compared to Qwen3's proprietary 36T-token corpus for pretraining, MobileLLM-R1-950M matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. To facilitate further research in this direction, we have released the complete training recipe, data sources, data mixing ratio, and model checkpoints, together with the key insights obtained throughout this study.
BibTeX
@inproceedings{zhao2026mobilellm,
  author = {Changsheng Zhao and Ernie Chang and Zechun Liu and Chia-Jung Chang and Wei Wen and Chen Lai and Sheng Cao and Yuandong Tian and Raghuraman Krishnamoorthi and Yangyang Shi and Vikas Chandra},
  title = {{MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=GMlZt4fZSY},
}
Figure