MobileLLM-R1: Sub-Billion Reasoning Models
- The paper introduces MobileLLM-R1, a dense decoder-only Transformer family that achieves competitive reasoning with sub-billion parameters through open-data curation and staged training.
- It employs a three-phase training process—pretraining, mid-training, and post-training—with capability-aware resampling and influence-based data selection to optimize reasoning performance.
- Empirical results show that the largest variant, MobileLLM-R1-950M, matches or exceeds benchmarks of larger models using only 11.7% of the tokens compared to proprietary pretraining corpora.
Searching arXiv for papers directly relevant to MobileLLM-R1 and closely related mobile/reasoning references. arXiv search query: "MobileLLM-R1" MobileLLM-R1 is a family of sub-billion-parameter “thinking” or reasoning LLMs trained from scratch with a fully open training recipe. Introduced in “MobileLLM-R1: Exploring the Limits of Sub-Billion LLM Reasoners with Open Training Recipes,” the project explicitly challenges the assumption that strong reasoning in compact models requires extremely large pretraining corpora, arguing instead that about $2$T tokens of high-quality open-source source data, resampled into a $4.2$T-token training exposure and followed by an established post-training procedure, are sufficient to produce competitive small reasoners (Zhao et al., 29 Sep 2025).
1. Definition and model family
MobileLLM-R1 is presented as a dense decoder-only Transformer family with three released scales: MobileLLM-R1-140M, MobileLLM-R1-360M, and MobileLLM-R1-950M. The architecture is described as being based on MobileLLM and LLaMA3.2, using the LLaMA3.2 tokenizer, a vocabulary of 128k subwords, QK-norm, and tied embedding/output weights. The training context lengths are staged: $2$k during pretraining, $4$k during mid-training, and $32$k during reasoning SFT; the family is presented as a $32$k-context model line (Zhao et al., 29 Sep 2025).
| Model | Core architecture | Parameters |
|---|---|---|
| MobileLLM-R1-140M | 15 layers, 9 heads, 3 KV heads, dim 576, FFN hidden dim 2048 | 140.2M |
| MobileLLM-R1-360M | 15 layers, 16 heads, 4 KV heads, dim 1024, FFN hidden dim 4096 | 359.4M |
| MobileLLM-R1-950M | 22 layers, 24 heads, 6 KV heads, dim 1536, FFN hidden dim 6144 | 949.2M |
The paper’s conceptual target is not generic mobile deployment alone, but the emergence of reasoning in compact models under open-data conditions. Two assumptions are placed under scrutiny: that reasoning only appears in sufficiently large models, and that such reasoning requires training on corpora larger than $10$T tokens. The first assumption is treated as already weakened by recent sub-billion reasoning models; the paper’s central intervention is to challenge the second by arguing that data quality, capability-aware curation, and mixture design matter more than raw corpus scale in this regime (Zhao et al., 29 Sep 2025).
A common misconception is to read the “R1” suffix as implying an RL-defined reasoning model. In this work, the main recipe is not an RL pipeline. The family is produced by staged pretraining, mid-training, and post-training, while GRPO appears only as an ablation rather than the defining method (Zhao et al., 29 Sep 2025).
2. Data curation and capability-aware pretraining
The data pipeline distinguishes sharply between the size of the curated source pool and the size of the actual training exposure. The underlying high-quality source inventory is reported as about $1.8$T tokens, described abstractly as “T,” while the model is trained on $4.0$T pretraining tokens plus $4.2$0T mid-training tokens, yielding $4.2$1T total sampled training tokens before or including reasoning SFT (Zhao et al., 29 Sep 2025).
The pretraining source datasets are StarCoder, OpenWebMath, FineWeb-Edu, Wiki, Arxiv, StackExchange, Algebraic Stack, Nemotron science, Nemotron code, Nemotron math, Cosmopedia, Facebook Natural Reasoning, FineMath, and peS2o. Their reported total is $4.2$2T tokens. Phase 1 pretraining assigns most probability mass to FineWeb-Edu at $4.2$3, followed by StarCoder at $4.2$4, OpenWebMath at $4.2$5, Arxiv at $4.2$6, StackExchange at $4.2$7, Wiki at $4.2$8, and Algebraic Stack at $4.2$9. Phase 2 shifts toward reasoning-heavy data: FineWeb-Edu remains largest at $2$0, but OpenWebMath rises to $2$1, FineMath to $2$2, Facebook Natural Reasoning to $2$3, and Nemotron math to $2$4, with smaller allocations to Cosmopedia, Arxiv, Algebraic Stack, StackExchange, Nemotron code, StarCoder, Wiki, peS2o, and Nemotron science (Zhao et al., 29 Sep 2025).
Mid-training consists of two phases of $2$5B tokens each. Phase 1 uses a Dolmino-centered mix led by Dolmino DCLM Baseline at $2$6 and Nemotron-Pretraining-Code-v1 at $2$7, while Phase 2 becomes even more concentrated on Dolmino Math at $2$8, Nemotron-Pretraining-Code-v1 at $2$9, and Nemotron-CC-Math-v1 at $4$0. A small “Benchmark Set” appears in mid-training phase 2 at $4$1, comprising training portions of TriviaQA, OBQA, NaturalQuestions, PIQA, GSM8K, BoolQ, ARC-Easy, and ARC-Challenge (Zhao et al., 29 Sep 2025).
The most distinctive part of the recipe is the benchmark-free, self-evolving data optimization framework. The authors construct capability-probing datasets for code $4$2, math $4$3, and knowledge $4$4. Representative subsets are obtained through a multi-stage procedure: a FineWeb-Edu classifier filter that keeps only samples with educational-quality score $4$5; Ask-LLM relevance scoring with a binary inclusion prompt and retention of the top $4$6 within each dataset; domain-specific Ask-LLM prompts; semantic deduplication; and compression of each corpus to about $4$7 representative examples. Dataset usefulness is then estimated by a leave-one-out loss delta,
$4$8
where positive $4$9 indicates that removing dataset $32$0 hurts capability performance (Zhao et al., 29 Sep 2025).
Sample-level weighting is based on an influence approximation,
$32$1
aggregated across code, math, and knowledge checkpoints:
$32$2
with $32$3 evenly spaced checkpoints and $32$4, so later checkpoints are weighted more heavily. Dataset-level sampling weights are then computed as
$32$5
During mid-training, the data-model co-evolution stage keeps only positive-influence samples,
$32$6
and iterates until most samples have near-zero or negative influence, which the paper interprets as information exhaustion (Zhao et al., 29 Sep 2025).
3. Training stages and post-training design
The training recipe has three major phases: pretraining, mid-training, and post-training. Pretraining is divided into two $32$7T-token phases. Each phase uses batch size $32$8, sequence length $32$9k, $32$0k steps, peak learning rate $32$1, hardware listed as $32$2 GPUs, and a wall-clock duration of $32$3–$32$4 days. Mid-training is also split into two phases, each with $32$5B sampled tokens, batch size $32$6, sequence length $32$7k, $32$8k steps, learning rate $32$9, the same $10$0 GPU configuration, and $10$1–$10$2 days per phase (Zhao et al., 29 Sep 2025).
The optimizer throughout pretraining is Adam with $10$3, weight decay $10$4, a $10$5k-step warmup, and linear decay to $10$6 peak learning rate. Mid-training also uses Adam, with learning rate decaying linearly to zero. Post-training again uses Adam but with zero weight decay; the warmup ratio is $10$7 for general SFT and $10$8 for reasoning SFT, followed by linear decay to zero (Zhao et al., 29 Sep 2025).
Mid-training incorporates knowledge distillation from a Llama-3.1-8B-Instruct teacher. The student is trained to minimize KL divergence between teacher and student logits. The paper does not provide the exact KL expression in the main text excerpt, but distillation is treated as a standard component of the mid-training phase rather than a special reasoning-specific innovation (Zhao et al., 29 Sep 2025).
Post-training is explicitly two-stage. First, the model is aligned with general instruction data using Tulu3-SFT, comprising $10$9K samples. Second, it undergoes reasoning-oriented SFT using OpenMathReasoning ($1.8$0M samples), OpenScienceReasoning-2 ($1.8$1K), and OpenCodeReasoning-2 ($1.8$2M), for a total of $1.8$3M reasoning-SFT samples. General SFT uses batch size $1.8$4, sequence length $1.8$5k, $1.8$6 epochs, and learning rate $1.8$7; reasoning SFT uses batch size $1.8$8, sequence length $1.8$9k, 0 epochs, and learning rate 1 (Zhao et al., 29 Sep 2025).
The paper is unusually explicit about what is not part of the main recipe. Final results do not rely on PPO, DPO, rejection-sampling finetuning, verifier or reranker stages, RLAIF, self-training during post-training, tool use, or RL as a required stage. A GRPO ablation is included, using a 2-step TULU3 cold start, NuminaMath-TIR, learning rate 3, batch size 4 prompts, 5 generations per prompt, and KL coefficient 6. The reported outcome is that a well-pretrained small base model can improve under RL, but SFT-distilled reasoning data outperform direct RL, and adding RL after full SFT gives little or no gain and can hurt small models (Zhao et al., 29 Sep 2025).
4. Empirical performance and benchmark profile
The paper reports both base-model and final post-trained results. For the 950M base model, MobileLLM-R1-950M-base reaches 7 on MATH500, 8 on GSM8K, 9 on MBPP, $4.0$0 on HumanEval, $4.0$1 on CommonSense average, and $4.0$2 on MMLU. Relative to Qwen3-0.6B-Base, it is slightly lower on MATH500 and MMLU, slightly higher on GSM8K, about tied on MBPP, and much higher on HumanEval. Against fully open baselines such as OLMo-2-0425-1B and SmolLM2-1.7B, the 950M base model is clearly stronger on reasoning-heavy base metrics (Zhao et al., 29 Sep 2025).
The final post-trained 950M model is the paper’s headline system. It reaches $4.0$3 on MATH500, $4.0$4 on GSM8K, $4.0$5 on AIME’24, $4.0$6 on AIME’25, and $4.0$7 on LiveCodeBench v6. These numbers define the paper’s strongest comparative claims. On AIME’24, MobileLLM-R1-950M scores $4.0$8, compared with $4.0$9 for OLMo-2-0425-1B-Instruct, $4.2$00 for SmolLM2-1.7B-Instruct, and $4.2$01 for Qwen3-0.6B. On AIME’25 it scores $4.2$02, slightly below Qwen3-0.6B at $4.2$03 but massively above OLMo-2 and SmolLM2, both at $4.2$04. On MATH500 it scores $4.2$05, essentially matching or slightly exceeding Qwen3-0.6B at $4.2$06 and far surpassing OLMo-2 and SmolLM2, both at $4.2$07. On LiveCodeBench v6 it reaches $4.2$08, above Qwen3-0.6B at $4.2$09, OLMo-2 at $4.2$10, and SmolLM2 at $4.2$11. GSM8K is less favorable: MobileLLM-R1-950M at $4.2$12 remains well above SmolLM2 at $4.2$13 but below OLMo-2 at $4.2$14 and Qwen3-0.6B at $4.2$15 (Zhao et al., 29 Sep 2025).
The smaller variants show a steep scale effect. The post-trained MobileLLM-R1-360M reaches $4.2$16 on MATH500, $4.2$17 on GSM8K, and $4.2$18 on LiveCodeBench v6. The post-trained MobileLLM-R1-140M reaches $4.2$19, $4.2$20, and $4.2$21 on the same three benchmarks. The paper notes that the $4.2$22M model still surpasses several $4.2$23B models on LiveCodeBench, while the $4.2$24M model remains weak in absolute terms but above the smallest open baselines (Zhao et al., 29 Sep 2025).
A central comparative point is token efficiency relative to Qwen3. MobileLLM-R1-950M is reported as being trained on only $4.2$25 of the tokens used for Qwen3’s proprietary $4.2$26T-token pretraining corpus, yet it matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. This is the empirical basis for the paper’s argument that extreme corpus scale is not a prerequisite for reasoning emergence in compact models (Zhao et al., 29 Sep 2025).
5. Relation to earlier MobileLLM work and adjacent “R1” systems
MobileLLM-R1 is best understood as a successor in the MobileLLM line, but not as a direct continuation of the earlier MobileLLM objective. “MobileLLM: Optimizing Sub-billion Parameter LLMs for On-Device Use Cases” focused on sub-billion architectures for mobile inference, arguing for deep-and-thin decoder-only designs, embedding sharing, grouped-query attention, and immediate block-wise weight sharing in MobileLLM-LS. That earlier paper studied $4.2$27M and $4.2$28M models, emphasized runtime and memory constraints on phones, and did not define any model called MobileLLM-R1 (Liu et al., 2024). MobileLLM-R1 inherits the naming lineage and the compact-model scale, but redirects the research question toward reasoning emergence, open-data curation, and staged reasoning post-training rather than on-device architecture alone (Zhao et al., 29 Sep 2025).
It should also not be conflated with “Mobile-R1: Towards Interactive Reinforcement Learning for VLM-Based Mobile Agent via Task-Level Rewards,” which is a VLM-based mobile agent trained in an Android emulator with a three-stage pipeline of format finetuning, action-level GRPO, and multi-turn task-level GRPO. That system operates on screenshots, historical actions, and an atomic action space of eight mobile commands such as click(x, y), type(text), and terminate(status). Its official name is Mobile-R1, not MobileLLM-R1, and it addresses interactive mobile GUI control rather than compact language-model reasoning from scratch (Gu et al., 25 Jun 2025).
A further neighboring reference is “Tool-R1: Sample-Efficient Reinforcement Learning for Agentic Tool Use,” which frames tool use as a reinforcement-learning problem over Thought $4.2$29 Code $4.2$30 Observation trajectories and introduces a dynamic sample queue to reduce online tool-execution cost. Methodologically, Tool-R1 represents the R1-style RL tradition more directly than MobileLLM-R1 does. The contrast is instructive: Tool-R1 treats multi-step tool interaction as the core optimization target, whereas MobileLLM-R1 explicitly finds that, in the sub-billion regime, careful pretraining plus reasoning SFT is more effective than trying to induce reasoning primarily through RL (Zhang et al., 16 Sep 2025).
6. Limitations, interpretation, and significance
The paper’s notion of “reasoning” is pragmatic and benchmark-centered. Improved performance on reasoning-heavy benchmarks is treated as evidence of reasoning behavior, but the work does not claim to establish human-like reasoning in a stronger philosophical sense. The benchmark-free claim applies specifically to pretraining mixture construction; the broader pipeline is not entirely benchmark-isolated, because mid-training phase 2 includes training portions of public benchmark datasets, and the paper does not provide a formal contamination audit (Zhao et al., 29 Sep 2025).
The compute budget, while modest relative to some frontier pretraining efforts, is still substantial in absolute terms. The recipe uses $4.2$31T training tokens, multiple phases, $4.2$32 GPUs, and several days per phase. The paper also notes tradeoffs in broad capability retention: symbolic reasoning gains can reduce MMLU and factual knowledge retention, and this degradation appears stronger in smaller models. Its RL conclusion is deliberately narrow as well, since the GRPO result is based on a short ablation rather than a broad theorem about all reinforcement-learning approaches for small models (Zhao et al., 29 Sep 2025).
Several implementation details remain under-specified in the supplied text, including exact positional encoding configuration, normalization beyond QK-norm, activation function, total FLOPs, and a full contamination methodology. These omissions do not obscure the central contribution, but they do limit exact reproducibility at the level of low-level systems replication (Zhao et al., 29 Sep 2025).
Within those limits, MobileLLM-R1 is significant for three reasons. First, it provides a concrete open-data argument against the view that reasoning emergence in compact models depends on $4.2$33T-token corpora. Second, it shows that data-centric mixture optimization, influence-based resampling, and staged post-training can produce sub-billion models that are competitive with much less open or much larger baselines. Third, by narrowing the reasoning gap at $4.2$34M–$4.2$35M scale, it makes strong compact reasoners a more credible ingredient in future on-device and mobile systems, even though the paper itself is not primarily a deployment study (Zhao et al., 29 Sep 2025).