Edge & SLMs
Small Language Models (<8B parameters) designed to run on laptops or mobile devices. Essential for local-only features and low-latency tasks.
| Rank | Model | Price | Summary |
|---|---|---|---|
|
1
|
Free | The Pocket Reasoner. Microsoft's update brings 'Reasoning' class capabilities to the edge. It can solve GSM8K math problems locally on an iPhone 16 Pro, making it the smartest sub-5B model available. | |
|
2
|
Free | The Multimodal Edge. Google's latest open lightweight model allows for native image and audio input directly on-device. It is the go-to for building local vision agents for robotics and smart home devices. | |
|
3
|
Research/Open | The Battery Saver. Specifically architected for Snapdragon and Apple Silicon NPU pipelines. It prioritizes watt-per-token efficiency, enabling always-on background intelligence without draining mobile batteries. | |
|
4
|
Free / Apache 2.0 | The Dense Powerhouse. Released June 2025, this model punches way above its weight class (8B params). It is the preferred choice for local RAG applications where higher reasoning is needed than what Phi-4 offers. | |
|
5
|
Open Source | The Privacy Core. A family of models (1B to 3B) optimized strictly for Apple Silicon. While it lags in general knowledge, it is unbeatable for on-device summarization and personal context management within the Apple ecosystem. |
Just the Highlights
Phi-4 Mini (3.8B)
The Pocket Reasoner. Microsoft's update brings 'Reasoning' class capabilities to the edge. It can solve GSM8K math problems locally on an iPhone 16 Pro, making it the smartest sub-5B model available.
Gemma 3 4B
The Multimodal Edge. Google's latest open lightweight model allows for native image and audio input directly on-device. It is the go-to for building local vision agents for robotics and smart home devices.
MobileLLM-Pro (Meta)
The Battery Saver. Specifically architected for Snapdragon and Apple Silicon NPU pipelines. It prioritizes watt-per-token efficiency, enabling always-on background intelligence without draining mobile batteries.
Mistral Small 3.2
The Dense Powerhouse. Released June 2025, this model punches way above its weight class (8B params). It is the preferred choice for local RAG applications where higher reasoning is needed than what Phi-4 offers.
OpenELM 2 (Apple)
The Privacy Core. A family of models (1B to 3B) optimized strictly for Apple Silicon. While it lags in general knowledge, it is unbeatable for on-device summarization and personal context management within the Apple ecosystem.