Skip to content
Stork.AIstork.ai
AI Tool

Bonsai 27B Review

Bonsai 27B is a 27-billion-parameter large language model developed by PrismML, designed to run entirely on a smartphone, completely offline.

shipped Jul 17, 2026aifreemium
ai
Bonsai 27B — product screenshot

Why it matters

1Bonsai 27B is a 27.8-billion parameter multimodal language model based on Qwen3.6 27B.
2It is available in a 1-bit binary variant (approximately 3.9GB) and a 1.58-bit ternary variant (approximately 5.9GB).
3The 1-bit variant achieves 11 tokens/s on an iPhone 17 Pro and retains over 90% of full-precision benchmark performance.
4Bonsai 27B offers a 14x reduction in size and 5x less energy consumption compared to conventional 16-bit models.

About Bonsai 27B

Pricing Plans

Bonsai 27B (Ternary)
  • • 5.9GB optimized for laptops
  • • Multi-step reasoning
  • • Tool calling
  • • Agentic workflows
Bonsai 27B (1-bit)
  • • 3.9GB for mobile devices
Bonsai Image
Bonsai 1.7B
Bonsai 4B
Bonsai 8B

Leadership

Unknown
API Docs

Specs

API Docs

View Documentation →

GitHub

View Repository →

API Available

Yes, public API

overview

What is Bonsai 27B?

Bonsai 27B is a large language model tool developed by PrismML that enables developers, tinkerers, and device manufacturers to deploy advanced AI capabilities on-device. It is designed for high-performance, on-device artificial intelligence, running entirely offline on consumer hardware like smartphones and laptops. Bonsai 27B is a 27.8-billion parameter multimodal language model based on Qwen3.6 27B. Its core innovation lies in its extreme compression, allowing it to run locally on consumer hardware such as smartphones and laptops, a capability previously impractical for models of its size. This enables phone-local 27B reasoning, with the 1-bit variant (approximately 3.9GB) fitting within the memory budget of high-end mobile devices like the iPhone 17 Pro. For laptops, the ternary variant (5.9GB) supports full 27B reasoning, tool use, and agentic workflows with a large 262K-token context window, suitable for long-document analysis and code work. The model's on-device inference ensures data privacy and effective operation with intermittent or no internet connectivity, making it suitable for privacy-sensitive and offline applications. It is built for serious engineering work, including reasoning through multi-step workflows, writing and debugging code, planning, refactoring, and supporting autonomous agentic execution locally. Furthermore, its multimodal understanding allows it to accept both images and text inputs, processing screenshots, documents, and camera input directly on the device.

features

Key Features of Bonsai 27B

Bonsai 27B offers a suite of features designed for high-performance, on-device AI, emphasizing efficiency and privacy.

  • API available for integration into custom applications.
  • Advanced reasoning capabilities for complex problem-solving.
  • Robust coding capabilities, including writing and debugging.
  • Supports agentic workflows directly on-device.
  • Designed for deployment on smartphones and laptops.
  • Operates completely offline, ensuring data privacy and accessibility.
  • Multimodal understanding, accepting both image and text inputs.
  • 14x less memory consumption compared to full-precision models.
  • 8x faster performance on supported hardware.
  • 5x less energy consumption for local inference.

use cases

Who Should Use Bonsai 27B?

Bonsai 27B is engineered for a diverse range of users and applications requiring high-performance, private, and offline AI capabilities.

  • Developers & Tinkerers: For building and experimenting with advanced AI agents and applications on consumer hardware.
  • Device Manufacturers: For integrating sophisticated AI capabilities directly into smartphones, wearables, and other edge devices.
  • Engineers in Robotics: For developing robot navigation subroutines and other on-device intelligence.
  • Engineers in Industrial Edge Applications: For visual inspection in manufacturing, keyword spotting in voice interfaces, and predictive maintenance on industrial equipment.
  • Users requiring Privacy-Sensitive Applications: For any use case where data must remain local and offline, such as advanced reasoning, coding, and agentic workflows on-device.

how to use

How to Use Bonsai 27B

Bonsai 27B models are available for download and can be integrated into various development environments for on-device execution.

  • 1Access the Bonsai 27B models, available under the Apache 2.0 license, from PrismML's official channels.
  • 2Download the desired variant (1-bit binary or 1.58-bit ternary) from Hugging Face repositories, which include GGUF and MLX formats.
  • 3Utilize the provided API for integrating Bonsai 27B into custom applications and workflows.
  • 4Refer to the PrismML documentation at https://prismml.com/docs/bonsai/run-the-server for server setup and deployment instructions.
  • 5Deploy the model on compatible consumer hardware, such as high-end smartphones (e.g., iPhone 17 Pro) or laptops.

pricing

Bonsai 27B Pricing & Plans

Bonsai 27B operates on a freemium model, with specific pricing details for its various models and tiers available upon direct inquiry to PrismML sales.

  • Bonsai 27B (Ternary): Contact sales
  • Bonsai 27B (1-bit): Contact sales
  • Bonsai Image: Contact sales
  • Bonsai 1.7B: Contact sales
  • Bonsai 4B: Contact sales
  • Bonsai 8B: Contact sales

Pros

  • +First 27B-class multimodal model capable of running on a smartphone.
  • +Extreme compression (1-bit and 1.58-bit quantization) results in 14x less memory consumption.
  • +High performance retention (over 90% for 1-bit, 95% for ternary) despite significant compression.
  • +Operates entirely offline, ensuring data privacy and functionality without internet.
  • +Significantly more energy-efficient for local inference (0.275 mWh/token on M5 Pro).
  • +Open-source availability under Apache 2.0 license, with Hugging Face integration.

Cons

  • −Performance gap concentrated in the most demanding categories, despite high retention.
  • −Agentic coding for long-horizon, multi-file workflows is not yet a strong target for this release.
  • −Initial compatibility issues reported with some third-party tools like LM Studio, indicating evolving tooling support.
  • −Specific pricing details for commercial use require direct contact with PrismML sales.
  • −Requires high-end mobile devices (e.g., iPhone 17 Pro) for optimal smartphone performance.

Similar Tools

Bonsai 27B vs Competitors

Bonsai 27B distinguishes itself in the on-device AI landscape through its 'Intelligence Density' philosophy, delivering high capability per unit of size, memory, power, and deployment footprint, particularly for a 27-billion-parameter model.

1
MLC Chat↗

MLC Chat is an open-source universal chat app that allows users to run various large language models directly on their devices, including smartphones, completely offline.

While Bonsai 27B is a specific 27-billion-parameter model, MLC Chat provides a platform to run a range of models, typically smaller (e.g., Llama 3.2, Gemma 2, Phi 3.5, Qwen 2.5, often in the 1B-8B parameter range), on-device and offline. It offers flexibility in model choice, whereas Bonsai 27B is a fixed model from PrismML.

2
PocketPal AIOn StorkCompare

PocketPal AI is a free, open-source application for Android and iOS that enables users to download and run GGUF-formatted LLMs entirely offline on their smartphones.

Similar to Bonsai 27B, PocketPal AI focuses on offline, on-device LLM inference for smartphones. However, it supports a variety of smaller GGUF models (e.g., Gemma 3 1B, Llama 3.2 1B/3B, Qwen2.5 1.5B), generally in the 1B-4B parameter range, rather than a single 27B model. It emphasizes user choice of models and privacy through local execution.

3
Google AI Edge Gallery / LLM Inference APIOn StorkCompare

Google AI Edge Gallery is an open-source Android application and platform showcasing the LLM Inference API, which allows developers to deploy and run various Google-optimized LLMs (like Gemma) directly on Android devices for offline use.

Bonsai 27B is a consumer-ready model, while Google AI Edge Gallery and its LLM Inference API are primarily developer-focused tools for integrating on-device AI into Android applications. Google's offerings typically feature smaller, optimized models (e.g., Gemma 2B, Gemma 3n) designed for mobile hardware, which are generally less than Bonsai's stated 27 billion parameters.

4
Apple Intelligence↗

Apple Intelligence is a suite of on-device AI capabilities deeply integrated into iOS, iPadOS, and macOS, processing most requests locally and offline using a 3-billion-parameter LLM.

Unlike Bonsai 27B, which is a standalone LLM, Apple Intelligence is a comprehensive system-level integration of AI features. It runs a smaller, 3-billion-parameter LLM on-device, prioritizing privacy and seamless user experience within the Apple ecosystem, whereas Bonsai 27B offers a significantly larger parameter count for a single model.

See every Bonsai 27B alternative, compared→
Visit Bonsai 27B↗
Connect
𝕏
X / Twitterx.com/PrismML
⌘
GitHubgithub.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf
in
LinkedInwww.linkedin.com/company/prismml/

AI Reputation Report

Is Bonsai 27B yours?

ChatGPT, Perplexity, Gemini, Claude & Grok answer buyer questions about Bonsai 27B every day. See whether they name Bonsai 27B — or send buyers to a rival.

See what AI saysfree preview

Featured in articles