PrismML has successfully compressed a 27-billion-parameter reasoning model into a 4GB footprint. This breakthrough could eliminate the need for cloud-dependent LLM deployment on mobile devices.
Large language model (LLM) deployment has long been constrained by a fundamental tension: the most capable reasoning models require data center-scale hardware, while the devices people actually carry in their pockets top out at a few gigabytes of usable memory. PrismML may have just broken that constraint wide open.
The AI infrastructure startup has compressed Bonsai 27B — a full 27-billion-parameter reasoning model — to under 4 GB, while retaining 90 percent of its original performance. The result is a reasoning-capable LLM that fits on an iPhone, no cloud connection required.
What PrismML Actually Built
Bonsai 27B is not a distilled mini-model or a stripped-down chatbot. It is a complete open reasoning model — the kind that can work through multi-step problems, generate chain-of-thought outputs, and handle complex inference tasks — compressed to a footprint that consumer mobile hardware can accommodate.
The 4 GB threshold is significant for a specific reason: it maps directly to the memory envelope available on current iPhone hardware after the operating system and active applications claim their share. Fitting a 27B-parameter model into that window required aggressive compression without gutting the model's reasoning capacity, which is where the 90 percent performance retention figure becomes the headline claim.
According to reporting by The Decoder, Apple is reportedly testing PrismML's compression technology — a signal that on-device AI inference at this scale is being evaluated at the highest levels of the mobile ecosystem, not just in research labs.
Why 90% Retention Under 4GB Is the Hard Part
Compressing large models is not new. Quantization, pruning, and knowledge distillation have been standard toolkit items for years. What has historically broken down is the performance cliff: aggressive compression tends to preserve surface-level fluency while degrading the deeper reasoning capabilities that make large models useful for enterprise tasks.
Reasoning models are particularly vulnerable because their value proposition depends on coherent multi-step inference chains. A model that loses 30 percent of its reasoning fidelity under compression isn't 70 percent as useful — for many enterprise workflows, it becomes effectively unusable for the tasks that justified the deployment in the first place.
PrismML's 90 percent retention claim, if it holds under independent evaluation, suggests the company has found a compression pathway that preserves reasoning structure rather than just surface outputs. The specifics of their methodology have not been fully disclosed, but the combination of sub-4 GB footprint with a 27B-parameter base model implies compression ratios that go well beyond standard INT4 quantization alone.
The Enterprise Mobile Deployment Angle
For technology decision-makers, the implications extend beyond the novelty of running a large model on a phone. Cloud-dependent AI inference carries real costs and constraints for enterprise mobile applications: latency on unreliable connections, data privacy concerns when sensitive information transits third-party infrastructure, per-token API costs that scale with usage, and hard dependencies on network availability.
On-device inference eliminates all four. A field service technician using an AI-assisted diagnostic tool doesn't need a 5G signal. A healthcare worker accessing a clinical decision support application doesn't need patient data leaving the device. A financial analyst running scenario models on a tablet in a low-connectivity environment gets the same performance as one sitting next to a fiber connection.
The shift from cloud-dependent to on-device reasoning isn't just a performance story — it's a deployment architecture story that changes how enterprises can build and distribute AI-powered applications.
Bonsai 27B, if it performs as described, represents a potential inflection point for this category. Previous on-device models capable of running on iPhone-class hardware have generally topped out at 7B or 13B parameters, with corresponding limitations on reasoning depth. A 27B model in the same footprint is a qualitative step up.
Apple's Reported Interest
Apple's reported testing of PrismML's compression technology adds a layer of strategic significance to the announcement. Apple has been methodical about on-device AI — its Neural Engine architecture, the Apple Intelligence framework introduced with iOS 18, and its on-device model infrastructure all reflect a long-term commitment to local inference as a privacy and performance differentiator.
If Apple is evaluating PrismML's approach, it suggests the compression methodology is credible enough to warrant serious engineering attention from a company that has built its own substantial in-house AI infrastructure. It also raises the possibility that future iPhone hardware or iOS releases could incorporate or build on similar compression techniques at the platform level.
What to Watch Next
Several questions will determine how significant this development ultimately proves to be:
Independent benchmarking of the 90 percent retention claim is the most critical near-term data point. Performance retention figures from model developers need third-party validation across diverse reasoning tasks — particularly the multi-step, chain-of-thought scenarios where compression tends to degrade most.
The open model question: Bonsai 27B is described as a full open reasoning model. How PrismML handles licensing, weights availability, and developer access will determine whether this remains a proprietary enterprise offering or becomes infrastructure that the broader developer community can build on.
Hardware generalization: The 4 GB target maps to iPhone, but the same compression techniques would apply to Android flagship devices, tablets, and edge computing hardware. If the methodology generalizes, the addressable deployment surface expands considerably.
Apple's next move: Whether Apple integrates, licenses, acquires, or simply monitors PrismML's technology will be a meaningful signal about where on-device reasoning sits in Apple's product roadmap for the next hardware cycle.
The era of cloud-dependent reasoning for mobile enterprise applications has been the default assumption for AI product teams for years. PrismML's Bonsai 27B is the clearest technical argument yet that this assumption deserves to be revisited.
Source: Bonsai 27B is a full open reasoning model that fits on an iPhone — The Decoder
Last reviewed: July 16, 2026
