Bonsai-27B is a Fully Open Reasoning Model That Runs on an iPhone
A new open-source AI reasoning model called Bonsai-27B fits entirely on an iPhone, offering full local inference without cloud dependency. The model packs 27 billion parameters into a size that runs directly on mobile devices, answering who (the research team), what (a fully open reasoning model), when (recently released), and why (to enable private, offline AI reasoning). Bonsai-27B is designed for tasks like math, logic, and step-by-step problem solving, all processed on-device.
The model demonstrates that large-scale reasoning capabilities can now work inside a smartphone.
Open Source and Transparent Weights
Bonsai-27B is released under an open license, with full model weights publicly available. This allows researchers and developers to inspect, modify, and deploy the model without restrictions. Unlike many commercial reasoning models, Bonsai-27B does not lock its architecture or training details behind proprietary walls.
The model was trained using openly documented methods and datasets.
Fit for Mobile Hardware
The key breakthrough is the model’s ability to run on an iPhone without needing a server connection. Standard 27-billion-parameter models require far more memory and compute than a phone can provide. Bonsai-27B achieves this by using advanced quantization and pruning techniques, shrinking the model footprint while retaining reasoning quality.
“This is the first fully open reasoning model that can be run entirely on a consumer mobile device.”
The team claims the model performs competitively with larger closed-source models on benchmarks like GSM8K and MATH. For real-world use, it allows privacy-sensitive applications like medical advice, tutoring, or personal assistants to keep all data on the device.
Performance Benchmarks
In testing, Bonsai-27B achieves strong scores on reasoning tasks:
- Math reasoning: Scores over 80% on GSM8K, matching models several times its size.
- Logic and deduction: Outperforms similarly sized open models on multi-step reasoning.
- Speed on iPhone: Inference takes only a few seconds per query, depending on device generation.
- Energy efficiency: Designed to avoid draining the battery during extended use.
These results show that open mobile AI can now handle complex cognitive tasks.
Implications for Privacy and Accessibility
Because the model runs entirely offline, user data never leaves the phone. This eliminates privacy risks tied to cloud AI services, where prompts and outputs may be stored or inspected. Developers can also integrate Bonsai-27B into apps without negotiating API licenses or paying per-token fees.
The open license ensures that anyone with an iPhone and basic technical skills can download and run the model. This lowers the barrier for researchers in regions with limited internet connectivity or restrictive data policies.
Technical Requirements
Running Bonsai-27B requires an iPhone with at least 6 GB of RAM and iOS 17 or later. The model uses Apple’s Core ML framework with optimized neural engine support. For developers, the model is available on Hugging Face and GitHub, along with scripts for quantization and deployment.
A desktop version for macOS and Linux is also planned, according to the project page.
Background and Development
The model was built by a team of researchers focused on making advanced AI accessible outside large cloud infrastructures. They started with a base language model and applied distillation and quantization techniques to reduce size without sacrificing reasoning depth. The training data was curated from publicly available sources, and the team documented all steps in an open research paper.
Bonsai-27B is part of a growing trend of small but powerful models that challenge the assumption that intelligence requires massive data centers.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.