The narrative for mobile AI has shifted. For years, the goal was simply shrinking massive server-grade models to fit on a phone. In late 2025, the focus is on architecture. New designs like Liquid Neural Networks, Mamba state-space models, and “Deep” Small Reasoning Models (SRMs) are outperforming traditional Transformers in battery life and logic, even on devices with strictly limited RAM (2GB–4GB).
Here is the current state of the art for on-device LLMs.
1. The Battery Efficiency King: LFM 2 (Liquid AI)
700M – 2.6B Parameters | Hybrid Liquid Neural Network
Liquid AI has moved away from the standard Transformer architecture entirely. The LFM 2 utilizes a “Liquid” neural network design (combining convolutions and attention), which processes data differently than the static weights of older models.
- Intelligence: This architecture balances immediate context with long-term adaptive memory. It excels at maintaining coherent conversations over time without getting stuck in the repetitive loops often seen in small Transformers. It won’t write a novel, but it keeps a chat flow natural.
- Hardware Reality: This is the most battery-efficient model on the list. The 700M variant is usable on practically any smartphone, including older 3GB RAM devices, while the 2.6B variant rivals 8B dense models in quality while running significantly cooler.
- Get it here: Hugging Face (LiquidAI/LFM2-1.2B)
2. The Logic Engine: Qwen 3 (Alibaba)
0.6B – 4B Parameters | MoE / Dense Hybrid
Qwen 3 stands out by integrating a “Thinking” process (Chain of Thought) directly into small-scale models. Usually, reasoning requires massive parameter counts, but Qwen optimizes this for the edge.
- Intelligence: The 4B variant is a powerhouse for math and coding. It generates internal reasoning steps to solve problems that typically stump models under 7B parameters. The 0.6B variant is the fastest usable LLM available, clocking 60+ tokens per second—perfect for background tasks like notification summarization where speed matters more than nuance.
- Hardware Reality: The 4B model runs hot and requires an 8GB RAM phone to perform well. The 0.6B model is lightweight but can be too brief in its responses.
- Get it here: Hugging Face (Qwen/Qwen3-4B-Instruct)
3. Deep Logic on Low Specs: Baguettotron (Pleias)
321M Parameters | 80-Layer Small Reasoning Model
Baguettotron takes a different engineering approach: depth over width. While most small models are “wide” (more parameters per layer), this model is extremely “deep” (80 layers) with a tiny total footprint.
- Intelligence: This model specializes in step-by-step logic puzzles and classification. It trades general knowledge for reasoning depth. It will fail at creative writing or trivia, but it punches significantly above its weight class for strictly defined logical problems. It creates “Reasoning Traces” to show its work.
- Hardware Reality: Extremely lightweight. It uses less than 600MB of VRAM, meaning it runs on virtually any Android device, including entry-level phones with 2GB RAM.
- Get it here: Hugging Face (PleIAs/Baguettotron)
4. The Context Processor: Granite 4.0 H Micro (IBM)
3B Parameters | Hybrid Mamba-2 + Transformer
Standard Transformers struggle with long documents on mobile because memory usage explodes as the text gets longer. IBM’s Granite 4.0 H uses Mamba-2 (a State Space Model) layers to fix this bottleneck.
- Intelligence: This model is built for RAG (Retrieval Augmented Generation). You can feed it entire books or long PDF reports, and it processes them with high fidelity. It prioritizes factual accuracy over personality. It is less “chatty” than Qwen but more reliable for data extraction.
- Hardware Reality: It requires about 1.8 GB of RAM (at Q4 quantization). Note that you must use the H-Micro (Hybrid) version to get the efficiency benefits; the standard Micro version is a regular transformer.
- Get it here: Hugging Face (ibm-granite/granite-4.0-h-micro)
5. The Compatibility Standard: Llama 3.2 1B (Meta)
1.23B Parameters | Pruned Transformer
While the models above use exotic architectures, Llama 3.2 remains the baseline generalist. It uses a standard Pruned Transformer design, ensuring it works on almost every AI app available.
- Intelligence: It offers stable instruction following for simple tasks. However, it lacks the specialized reasoning depth of Qwen or Baguettotron and can struggle with nuance in complex prompts.
- Hardware Reality: It fits comfortably in ~1GB of RAM (Q4). It is the reliable fallback if the specialized architectures (Liquid/Mamba) aren’t supported by your specific mobile app.
- Get it here: Hugging Face (meta-llama/Llama-3.2-1B-Instruct)
