PrismML launches Bonsai 27B, a model based on Qwen3.6 27B that it says runs natively on Apple devices via MLX; its CEO says Apple is evaluating the tech
Apple is in talks with a small Silicon Valley company that says it can shrink powerful artificial intelligence models enough to run directly …
CNBCMacKenzie Sigalos
Context & Ripple Effects
Apple’s on-device AI effort has previously centered on an approximately 3B-parameter language model, alongside larger models served through Private Cloud Compute. Its MLX framework, released in 2023, provides the Apple-silicon deployment path PrismML says Bonsai 27B uses.
Related reporting had already said PrismML ran a Qwen 3.6 27B model on an iPhone 17 Pro and that Apple had discussed the company’s technology. The launch turns that claimed demonstration into a named model offering while Apple reportedly evaluates the approach.
First-order effects
PrismML gains a concrete product reference for its Apple-device inference technology and a potential validation point as Apple evaluates it.
Apple has another potential route to run a substantially larger model locally through MLX, rather than limiting capable experiences to its smaller on-device model or server-backed processing.
Second-order effects
If the claimed deployment is reproducible, Apple’s internal model and platform teams will face pressure to compare compression, memory use, latency, and quality against their existing on-device and Private Cloud Compute split.
Developers using Apple silicon and MLX gain a stronger signal that larger local models may be practical, potentially increasing demand for tools that optimize models for constrained device memory and power.
Third-order effects
The episode points toward a more flexible device-cloud AI architecture: model size alone may become less decisive than the ability to package and run models efficiently on a given hardware stack.
Whether this changes Apple’s product strategy depends on independent performance and reliability validation; the durable industry shift would be toward deployment-specific model optimization rather than one model serving every endpoint.
The trend: AI deployment is shifting from simply scaling model parameter counts toward tailoring and compressing capable models for local execution on device-specific hardware.
Here is Ternary Bonsai 27B running an end-to-end agentic workflow locally with Hermes on an NVIDIA GeForce RTX 5090 GPU. The model reasons, calls tools, reads outputs, modifies files, and surfaces insights - all on consumer hardware, while all private files, intermediate states, …
The phone threshold is even harder than the storage number suggests. A phone exposes only part of its memory to an application, and the model must share that budget with its KV cache, activations, runtime, and the rest of the product. At 3.9 GB, 1-bit Bonsai 27B clears that [vide…
The footprint reduction does not come at the expense of the capabilities that matter. Across 15 benchmarks spanning knowledge, reasoning, math, coding, instruction following, tool use, and vision, Ternary Bonsai 27B retains 95% of the full-precision model's performance. The [imag…
Why does this matter? Because modern AI workflows are no longer single prompts. They are sustained loops. A capable agent may take hundreds of steps: reasoning, calling tools, reading outputs, updating its state, and iterating toward a result. When every step is remote, [image]
Raw capability determines what a model can do. Intelligence density determines where it can do it. Bonsai 27B moves the Pareto frontier left again: 27B-class capability in a footprint smaller than many full-precision 2B models. By intelligence density, 1-bit Bonsai 27B delivers […
Today, we're announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, […
‘The technology, according to PrismML, could ultimately extend well beyond phones and laptops to robotics, autonomous systems and other products that need to make decisions quickly without relying on a cloud connection.’ www.cnbc.com/2026/07/14/a...
it fits into 3.9 GiB!! — this is part of the new trend — tiny LLM that's part of a larger multi-agent system (involving frontier models, software & more) — an LLM small enough to keep in your pocket, to use obsessively for long-running agent tasks [embedded post]