The era when artificial intelligence required powerful servers and a constant internet connection may be coming to an end. Startup PrismML has unveiled a technology capable of reshaping the mobile industry landscape: they have created the Bonsai 27B model, which runs directly on smartphones, matching the capabilities of cloud giants.

The innovation is based on the open-source Qwen3.6 27B neural network from China's Alibaba. In its "native" form, it weighs about 54 GB, making it impossible to run on consumer gadgets. PrismML engineers applied a method of radical simplification of mathematical values: instead of standard 16-bit encoding, every weight inside the neural network is now recorded in just 1–2 bits. Patents for this method belong to the California Institute of Technology (Caltech) and are licensed exclusively to the startup, which raised $16.25 million in investment in March.

Technological breakthrough and its price

The company has released two free versions of the model under the Apache 2.0 license. According to developers, such extreme compression allows the model to use 10–15 times less RAM. At the same time, the speed of generating responses increases by 6–8 times, while energy consumption decreases by 3–6 times.

However, physics does not forgive miracles, and the technology has its compromises. Extreme compression leads to a loss of a few percent of overall efficiency. The Bonsai 27B model shows noticeably worsened abilities in citing exact facts, image recognition, and working with external tools. Nevertheless, logic, mathematics, and code writing remain at a high level.

Also, the question of energy consumption during prolonged use remains open. Continuous operation of the neural network in the background can quickly drain the battery, and the processor begins to throttle frequencies due to overheating after just five minutes of continuous operation.

Apple's strategy: why iPhone is interested in Bonsai

Apple is holding active negotiations with PrismML against the backdrop of the launch of the public beta version of iOS 27, where the company presented a major update to Siri. Local launch of compressed models directly on iPhone processors gives the "Apple" corporation several strategic advantages:

  • Minimal latency: the smartphone no longer needs to wait for a response from the cloud — request processing happens instantly.
  • Maximum privacy: sensitive data, such as medical information, medication lists, or fitness metrics, never leave the device.
  • Autonomy: basic AI functions, computational photography tools, video generation, and text editing can work without an internet connection.
  • Cost reduction: reduces the financial burden of maintaining Apple's own data centers and paying for external cloud licenses.

Analysts' opinion

Despite PrismML's optimistic statements, independent analysts from Counterpoint Research and IDC urge caution in evaluating the technology. Real-world tests outside the laboratory have not yet been conducted, and it is too early to say whether this technology can become an industry standard.