Back to News
ProductAI Understanding briefing

PrismML releases Bonsai 2, a compressed 27B model for local devices

PrismML has released Bonsai 2, a 5.9GB compressed version of Alibaba's Qwen3.8 27B model that runs locally on smartphones and PCs while retaining 98% of the original's benchmark performance.

4 min readRead the linked source
Source-provided image accompanying PrismML releases Bonsai 2, a compressed 27B model for local devices
Source referenceSource recorded
Publisher
techjuice.pk
Source link
techjuice.pkhttps://www.techjuice.pk/prismml-bonsai-2-27b-llm-compression-smartphone-pc-ai-device/
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Start here

Key terms

Large Language Model (LLM)
A language model trained on massive text corpora to generate and analyze text.
On-Device AI
AI inference performed locally on user hardware rather than in a remote cloud service.
Benchmark
A standardized test or dataset used to measure and compare model performance.
Test yourselfAI Models Explained Quiz

What happened

PrismML released Bonsai 2 27B, a compressed large language model designed to run on personal computers and high-end smartphones without cloud infrastructure. The model reduces Alibaba’s Qwen3.8 27B to 5.9 gigabytes using ternary weight compression, retaining 98% of the original’s performance. The weights are available under the Apache 2.0 License and support NVIDIA GPUs via CUDA and Apple devices via MLX.

PrismML, a startup founded by Caltech researchers and led by Babak Hassibi, released Bonsai 2 27B on Thursday. The model is a compressed version of Alibaba’s Qwen3.8 27B, reduced to 5.9 gigabytes through ternary weight compression, which converts 16-bit parameters to three values: positive one, negative one, or zero. This technique eliminates redundant parameters, allowing the model to retain 98% of the original’s performance while fitting on local hardware.

The release marks a significant improvement over the original Bonsai model from March, which retained 95% of performance. According to PrismML, the original model has been downloaded over 11 million times, with additional smaller models accumulating 2.6 million downloads, indicating strong developer adoption for device-based AI. The new model runs on NVIDIA GPUs via CUDA and on Apple devices (Mac, iPhone, iPad) via MLX, utilizing custom low-bit kernels to ensure efficient local execution.

The model weights are available immediately under the Apache 2.0 License, allowing for broad integration and modification. PrismML raised a $22.25 million seed round backed by Khosla Ventures, Cerberus Capital, and Caltech, with Ion Stoica, Databricks co-founder, serving as an adviser. The company’s approach focuses on leveraging the redundancy in larger models to achieve high performance parity with minimal file size, a strategy Hassibi suggests will become more effective as model sizes increase.

Source details: techjuice.pk

Why it matters

This release demonstrates that reasoning-capable AI can operate locally on consumer hardware, addressing privacy concerns and eliminating cloud latency and subscription costs. By achieving high performance parity with a significantly smaller file size, PrismML advances the shift from cloud-dependent AI to distributed, device-based reasoning, making advanced AI capabilities accessible without massive data center infrastructure.

The ability to run reasoning-capable AI locally on consumer devices addresses critical privacy concerns by ensuring user data never leaves the device. This eliminates the need for cloud transmission, reducing latency and removing ongoing subscription costs associated with cloud-based AI services. For developers and end-users, this represents a fundamental shift toward distributed, device-based reasoning that is both cost-effective and private.

The high retention rate of 98% performance in a 5.9GB file demonstrates that compression techniques can effectively preserve the intelligence of large models. This has practical implications for industries requiring , such as healthcare, finance, and personal assistants, where data sovereignty and real-time response are paramount. The open-source nature of the release further accelerates adoption by allowing the community to build upon and optimize the model for specific hardware constraints.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
AI Models Explained Quiz

What is the best response when AI Models Explained makes a mistake in production?

What to watch next

Developers should monitor the availability of custom low-bit kernels for different hardware architectures and the company's stated goal of compressing hundreds-of-billions-parameter models in the coming months. Additionally, the potential integration with Apple devices, hinted at by the timing of the release, warrants attention for future device-side AI capabilities.

PrismML has stated its next goal is to compress models with hundreds of billions of parameters within the next couple of months. Success in this area would further validate the scalability of ternary compression for frontier-class models, potentially making even more advanced AI capabilities available on local hardware.

The timing of the release has led to speculation about potential integration with Apple devices. While CEO Babak Hassibi declined to comment on reports of talks with Apple, the support for Apple’s MLX framework suggests a strategic alignment with Apple’s ecosystem. Future announcements regarding official partnerships or device-specific optimizations could significantly impact the adoption of local AI on iOS and macOS platforms.

Related guides & quizzes

AI Models ExplainedAI TrainingAI AgentsTest what you know — try a free AI quizLook up an AI term in our glossary
Found this useful?