Today, Oscilla is introducing OSLO 1.5B, our first fine-tuned language model and the beginning of a broader effort to explore how artificial intelligence can be developed specifically for the devices people already own.

Built on Qwen 2.5, OSLO 1.5B is a compact language model designed for responsive, everyday use on consumer hardware. With 1.5 billion parameters, a 32K context window, and an on-device footprint of approximately 1.63 GB, OSLO is designed to run locally on iPhone 13 and newer, bringing capable AI to devices already in people's hands.

This release represents an important expansion of Oscilla's work in local AI. Until now, our focus has been on building the software that makes it easier to run and interact with open models across Apple devices. With OSLO, we are extending that work into model development itself, allowing us to better understand, optimize, and build the intelligence that powers the experiences we create.

Why we're building models

Local AI is moving beyond a technical experiment toward a practical way of computing. Consumer hardware continues to improve, open models are becoming smaller and more capable, and the gap between what can run in the cloud and what can run locally continues to narrow. This shift creates an opportunity to rethink how AI models are developed, not simply where they run.

Most language models are developed with cloud infrastructure in mind, where large amounts of memory and computing power are readily available. Running those same models locally introduces a different set of constraints. Memory consumption, inference speed, battery life, thermal performance, context length, and quantization become central considerations rather than secondary optimizations.

Developing OSLO allows us to study those tradeoffs directly and begin shaping model behavior around the realities of local hardware. Instead of treating a model as a fixed component, we can explore how its architecture, training, and optimization should evolve when an iPhone or Mac is the intended environment from the beginning.

Our goal is not to replicate the largest cloud models at a fraction of their size. It is to better understand what makes a model genuinely useful when privacy, efficiency, responsiveness, and accessibility are fundamental requirements.

Designed for the local environment

OSLO 1.5B was intentionally developed as a lightweight model. At 1.5 billion parameters and approximately 1.63 GB, it is designed to make local inference practical on hardware that millions of people already own. Support for a 32K context window also allows the model to work with longer conversations and larger amounts of information while maintaining a compact footprint.

Starting with a smaller model is a deliberate choice. We believe the future of local AI will depend not only on increasingly capable models, but on making those capabilities available across a wide range of consumer devices. Supporting iPhone 13 and newer gives us a broad hardware baseline for understanding real-world performance and improving how models operate across different generations of hardware.

Model size, however, is only one part of the equation. A compact model can become considerably more useful when integrated into a broader system that provides capabilities beyond its underlying weights. Within Oscilla, models can work alongside persistent memory, user-provided files, personal context, voice, and device integrations to support experiences that would otherwise require a much larger model.

This relationship between the model and its surrounding environment is central to our approach. We believe the quality of on-device AI will increasingly depend on how effectively models, software, and hardware work together, rather than on parameter count alone.

Building the full local AI stack

Oscilla was founded on the belief that people should be able to use capable artificial intelligence without giving up control of their data or relying entirely on remote infrastructure. That principle has guided our development of tools that allow people to run open models directly on their devices, interact with local information, and create personalized AI experiences.

OSLO extends that mission beyond the application layer and into the models themselves.

Working with existing open models has given us valuable insight into how AI performs once it reaches a device. Developing our own fine-tuned models provides a deeper understanding of the processes that shape that performance, including fine-tuning, instruction following, quantization, context management, memory requirements, and tool use.

More importantly, it creates a closer development loop between the model, the hardware, and the application. We can evaluate where OSLO performs well, identify its limitations, explore how different configurations affect performance, and improve the software surrounding it in response. Changes to the model can inform product development, while insights from the product can guide future model improvements.

This approach also gives us greater flexibility to investigate capabilities designed specifically for local use. Rather than adapting every experience to the behavior of a third-party model, we can begin developing models with a clearer understanding of the tasks, integrations, and constraints they are intended to support.

Over time, we expect that relationship to play an increasingly important role in how Oscilla evolves across iOS and macOS.

The beginning of the OSLO family

OSLO 1.5B is our first step toward what we intend to become a broader family of models developed with local computing in mind. We do not believe a single model will be ideal for every device or use case. Different hardware has different capabilities, and different tasks benefit from different approaches to model size, architecture, and specialization.

Future development may explore additional model sizes, multimodal capabilities, stronger tool use, more efficient quantization, improved context handling, and models optimized for particular devices or tasks. Each of these areas presents an opportunity to better understand how intelligence can operate efficiently and effectively without depending on cloud infrastructure.

The initial release is as much a research milestone as it is a product milestone. We plan to continue evaluating OSLO across supported hardware, studying its performance, and using those findings to improve future iterations. The lessons we learn will also influence the broader Oscilla platform, from inference and memory management to how models interact with tools, files, and other device capabilities.

What comes next

With OSLO, we are beginning a longer effort to build AI around the capabilities of personal devices rather than the infrastructure of remote data centers. Our focus remains on making local intelligence more capable, accessible, and useful while preserving the privacy and control that make on-device computing valuable.

There is still a great deal to learn about how models should be designed for this environment. OSLO 1.5B gives us a foundation to explore those questions directly and apply what we discover across the entire Oscilla experience.

Until now, we have focused on building the software that brings local AI to people. With OSLO, we are beginning to build the intelligence behind it.

This is the beginning of OSLO, and the next chapter of Oscilla.