ArticleNews
Running high-parameter language models directly on mobile hardware has long been blocked by memory constraints. Compressing a 27-billion-parameter model below 4 GB brings frontier-level capabilities onto mobile devices, reducing server infrastructure overheads and improving user privacy.
Deploying large models on-device is central to Apple's AI economics: running processing locally lowers cloud compute expenses and strengthens privacy guarantees. PrismML's reported capability to keep all 27 billion parameters active sets it apart from common sparse models that selectively activate weights. If independent tests confirm the performance claims, such mathematical compression techniques could accelerate the shift from cloud-dependent assistants to standalone edge AI.
Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.
Similar in content, found by the AI search.
Article:Joe Rossignol
Apple has held discussions with startup PrismML to evaluate running substantially larger AI models directly on iPhones, according to a report by The Information. PrismML reportedly compressed Alibaba's 27-billion-parameter open-source model Qwen 3.6 to run entirely on an iPhone 17 Pro.
Link:prismml.com
PrismML has introduced the Bonsai model family, offering ultra-dense AI models based on research from Caltech. The flagship Bonsai 27B model runs locally on smartphones in a 1-bit variant requiring only 3.9 GB of memory.
X post:Hamid Maei
PrismML has announced Ternary Bonsai 2 27B, a compressed version of Qwen3.8 27B. It is nine times smaller than the full-precision original, requiring only 5.9 GB of storage while retaining 98.2% of aggregate benchmark performance.
97bookmarks21kviews
X post:Udi Wertheimer
Udi Wertheimer highlights PrismML's release of Ternary Bonsai 2 27B, a model based on Qwen3.8 27B compressed to 5.9 GB while reportedly retaining 98.2% of its full-precision benchmark performance.
1.5kbookmarks260.1kviews
Article:Zhidx
Ant Group's Bailian team has released Ling-3.0-tiny, an open-source lightweight mixture-of-experts model. It leverages DeepSeek's MLA and Kimi's KDA architectures with 7.9B total parameters, activating just 1.3B parameters per token for local edge deployment on hardware like Apple Silicon Macs and DGX systems.
Link:PocketAiHub
PocketAiHub has released an unofficial MLX derivative family of Qwen3.8-27B with suppressed refusal behaviour ('abliteration'). The consolidated repository provides quantized builds ranging from experimental 2-bit AWQ to full BF16 optimised for Apple Silicon.
Parts of this content were created with the assistance of AI.
