A Caltech spinout called PrismML released a compressed 27-billion-parameter AI model on July 14 that fits in under 4 GB of memory and runs on any iPhone 15 or newer — and Apple is actively testing whether the underlying technology could change how the company handles on-device AI. PrismML CEO Babak Hassibi confirmed to CNBC that Apple and unnamed other companies have been evaluating the startup’s models, measuring speed, energy efficiency, and device-level performance. “They’re really evaluating our technology right now,” Hassibi said of Apple, characterizing the discussions as early-stage but adding that “things are progressing nicely.” Apple did not respond to requests for comment.
The timing is precise. Apple opened the first public beta of iOS 27 on July 13 — one day before PrismML released its model — as CNBC noted in its reporting on the talks. The new iOS revision is built around Siri AI — a rebuilt personal assistant that Apple depends on a hybrid of on-device and cloud infrastructure to power. Whether PrismML’s compression technique could reduce the cloud portion of that equation is exactly what the evaluation is probing.
How 27 Billion Parameters Fit Into 4 GB
The scale of the compression is the story. Standard large language models store each internal weight as a 16-bit floating-point number. Alibaba’s Qwen3.6 27B — the open-source base model PrismML worked with — occupies roughly 54 GB at that precision, which puts it far beyond the memory budget of any smartphone. Even an aggressive 4-bit quantization build of the same model still requires approximately 18 GB, making it unusable on a phone and impractical on most laptops.
PrismML’s Bonsai 27B takes a more radical approach: it reduces each weight to a single bit in its binary variant — the value is either +1 or -1, nothing else — or to one of three values ({-1, 0, +1}) in its ternary variant, which requires a theoretical 1.58 bits per weight. This compression is applied end-to-end: embeddings, attention layers, the multi-layer perceptron blocks, and the language model head all use the low-bit format. There are no higher-precision escape hatches left in the architecture.
To keep the model accurate despite such extreme compression, the company uses a technique called FP16 group-wise scaling: each cluster of weights retains its own full-precision (16-bit) scaling factor, anchoring the model’s outputs even as individual weight values collapse to discrete steps. The result, in PrismML’s own benchmarks, is a 1-bit binary model at 3.9 GB that retains more than 90% of the base model’s performance across 15 combined evaluations, and a 1.58-bit ternary model at 5.9 GB that retains more than 95%. Both figures are from the company’s own testing; independent third-party verification has not yet been completed.
On an iPhone 17 Pro, the 1-bit model runs at 11 tokens per second. On an M5 Max, it reaches 87 tokens per second in its 1-bit variant. By the company’s own intelligence-density metric — capability per gigabyte of deployment footprint — the 1-bit Bonsai 27B delivers 0.53 per GB, which it describes as more than 10 times the full-precision baseline and roughly 2.7 times the best conventional low-bit alternative at the same parameter count.
The acknowledged tradeoff, which CEO Hassibi disclosed to CNBC, falls primarily on factual recall: that capability degrades before higher-order skills like reasoning, mathematics, and coding. Agentic task performance — tool use, multi-step planning, coding — holds up comparatively well in PrismML’s reported figures.
Sparse Activation vs. Dense Compression: Two Different Engineering Bets
The technical distinction between PrismML’s approach and Apple’s current on-device AI architecture is not aesthetic — it reflects a divergent answer to the same underlying engineering problem.
Apple’s AFM 3 Core Advanced, announced at WWDC 2026 on June 8, is a 20-billion-parameter model that uses a technique called Instruction-Following Pruning (IFP). The full 20 billion parameters are stored in flash (NAND) memory. For each request, the model makes a routing decision and loads only a subset — between 1 and 4 billion parameters — into active DRAM. The rest stays in flash. Apple chose this architecture because the bandwidth between NAND and DRAM is too slow to swap parameters token-by-token (as standard mixture-of-experts models require), so the system commits to a specific parameter subset per prompt rather than per token.
This approach allows a very large model to fit on a phone without occupying the full DRAM budget — but the model is only sparsely active. The parameters not loaded for a given request contribute nothing to that inference.
PrismML’s approach keeps every parameter active at all times. The entire 27-billion-parameter model sits in DRAM simultaneously at under 4 GB, because each parameter has been quantized to a single bit. There is no flash-to-DRAM routing bottleneck and no parameter subset selection. The cost is that all 27 billion weights have been reduced to three possible values instead of 65,536.
There is a hardware implication that neither company’s announcement explicitly addresses: native chip support is required to translate ternary weights into actual inference speed gains. Standard CPUs and GPUs must dequantize ternary weights back to a higher precision to use their existing multiply-accumulate hardware units, which can negate the computational benefit of the compression. Whether Apple’s Neural Engine, which is designed for fixed-point arithmetic, can execute 1-bit matrix operations natively without this overhead is an open question — and one with direct implications for how much engineering Apple would need to do if it were to adopt PrismML’s models for Apple Intelligence.
Why Apple’s Internal Compression Attempt Reportedly Failed
Multiple independent reports, citing anonymous sources, indicate that Apple encountered significant performance degradation when it attempted internally to compress its own models to run on iPhones. The specific models involved and the degree of degradation are not publicly known, but the reports suggest the failure is what opened the door to PrismML being evaluated externally.
Apple’s strategic incentive to solve this problem is quantifiable. Bank of America analysts, writing after the AFM 3 announcement at WWDC 2026, estimated that Apple’s most advanced cloud-based model, AFM 3 Cloud Pro, could account for roughly 5% of all Siri AI queries but up to 67% of weighted cloud-computing costs, because of the higher processing intensity required for complex reasoning and image-generation tasks — as reported by TheStreet. AFM 3 Cloud Pro runs on Nvidia GPU hardware inside Google’s cloud infrastructure — a dependency that sits uneasily with Apple’s privacy-first marketing and its competitive relationship with Google.
An on-device model capable of handling complex reasoning tasks would reduce the fraction of queries that flow to that expensive infrastructure. For a company projecting a record roughly $140 billion in free cash flow for 2026 while its AI cloud costs scale with every Siri interaction, that is a financially material engineering outcome.
Industry Caution: Real-World Performance Still Unproven
The benchmark figures that PrismML released are self-reported and have not been independently verified by third parties. Analysts who follow the mobile AI market have been direct about what that means in practice.
Tarun Pathak, research director at Counterpoint Research, said the controlled demonstration does not settle the real deployment questions. “The ultimate test will be millions of queries, thousands of device combinations and robust testing at scale,” Pathak told CNBC. Phil Solis, who leads client processor research at IDC, flagged battery life as potentially the defining variable: a model capable enough to be invoked frequently, or running persistently for agentic tasks, could drain a phone’s battery even if each individual inference uses less energy than its cloud equivalent.
Extreme quantization has a documented history of instability in specific domains. PrismML’s use of group-wise scaling factors improves stability compared to pure 1-bit quantization with no correction — research has confirmed that unmitigated 1-bit quantization degrades much faster on complex tasks. Whether the architecture holds up on adversarial prompts, very long context windows, or multilingual use at scale has not been tested by independent researchers with access to the model.
There is also a practical memory constraint the headline numbers do not capture. A phone’s useful memory budget for an AI model is narrower than its raw capacity. A 12 GB iPhone exposes approximately 6 GB of memory to an application — and the model must share that budget with the key-value cache and activations that grow with context length. The 3.9 GB figure for the 1-bit model leaves room to work within that constraint, but the gap is not as wide as the 12 GB vs. 4 GB comparison suggests.
One editorial observation from 9to5Mac is worth recording: the startup’s decision to publicly announce the Apple evaluation coincides directly with the Bonsai 27B launch, which is a standard startup publicity move. 9to5Mac noted that keeping serious Apple discussions quiet is generally the wiser course, but acknowledged that PrismML had succeeded in attracting attention from The Information and CNBC around its release — short-term mission accomplished.
What an Apple Partnership Could Mean for the Chip Market
If the Apple-PrismML evaluation leads to a formal relationship — a licensing agreement, a technology integration, or an acquisition — the downstream effects would extend beyond the iPhone product line.
At sufficient scale, on-device inference for complex AI tasks would reduce the volume of queries routed to cloud GPU infrastructure. The implications would be felt in demand for high-bandwidth memory chips and GPU server capacity in consumer-facing AI workloads — currently among the highest-growth segments in the AI data center market. Decentralized compute networks and cloud inference providers would face a shrinking addressable market in exactly the use cases that are easiest to capture on-device first: private, latency-sensitive, frequently repeated queries.
On-device AI at this capability level also changes what is economically possible on consumer hardware that is already deployed. PrismML’s Bonsai 27B model is available today, free, under an Apache 2.0 open-source license, on Hugging Face, compatible with Apple’s MLX framework and Nvidia’s CUDA. Developers, researchers, and Apple’s own engineering teams can test the real-world behavior of the model without waiting for any business relationship to be formalized.
CEO Hassibi has stated publicly that the company plans to compress larger frontier models next — a compressed version of Google’s Gemma model is in the pipeline, followed by work on models above 27 billion parameters. “Imagine, maybe three years from now, 95% of the intelligence you need is available locally — on your phone, laptop, home appliances — and only that last 5% of high-end needs actually requires the cloud,” Hassibi told CNBC. The company’s seed funding from Khosla Ventures, Cerberus Capital, and Caltech — a combined $16.25 million, modest by current AI standards — gives it runway to pursue that roadmap without near-term pressure to generate revenue.
Hassibi’s longer-term vision extends beyond smartphones. Robotics, autonomous systems, medical devices, and industrial sensors — environments that need fast, private AI decisions without a reliable cloud connection — are all potential deployment environments if the compression methodology scales to larger and more diverse model architectures.
Whether the Apple discussions lead anywhere remains genuinely unknown. But the model’s public release means the technology’s real-world capabilities will be tested at scale regardless of what Apple decides.
Frequently Asked Questions
What makes PrismML’s approach different from Apple’s own on-device AI?
Apple’s AFM 3 Core Advanced uses sparse activation: a 20-billion-parameter model stored in flash memory, with only 1 to 4 billion parameters loaded into active memory per request. PrismML’s Bonsai 27B uses extreme quantization: all 27 billion parameters are always active in memory simultaneously, but each parameter has been reduced to one or three discrete values instead of 16-bit floating-point. The two approaches address the same problem — fitting large models onto phones — through opposite engineering choices. Apple’s approach avoids the quantization accuracy tradeoff but creates a flash-to-memory bandwidth constraint. PrismML’s approach eliminates the bandwidth bottleneck but introduces an accuracy penalty and potentially requires native chip-level support to realize full speed gains.
Will running a 27-billion-parameter AI model drain an iPhone’s battery?
This is the key unanswered question. PrismML reports that its 1-bit Bonsai 27B uses 3 to 6 times less energy than a conventional equivalent running on existing hardware — but this is the company’s own benchmark figure and has not been verified by independent researchers. Phil Solis of IDC has specifically flagged battery life as the potentially defining variable: a model capable enough to be invoked frequently, or running as a background agent, could drain battery over time even if individual inferences are efficient. Independent testing at realistic usage levels has not yet been published.
Can Apple’s Neural Engine run 1-bit models without dequantizing the weights first?
This is the largest open technical question for any potential Apple-PrismML partnership. Standard CPUs and GPUs must dequantize ternary weights — converting them back to a higher-precision format — before using their multiply-accumulate hardware, which can negate the speed and energy benefits of the compression. Apple’s Neural Engine uses fixed-point arithmetic, which is more naturally compatible with low-bit computation, but whether it can execute 1-bit and ternary operations natively without a dequantization step depends on hardware specifics Apple has not publicly disclosed. If a chip redesign is required for full performance, the depth of engineering investment needed for an integration would be substantially larger than a software-only partnership.
Is Bonsai 27B available to download now?
Yes. PrismML released Bonsai 27B on July 14 under an Apache 2.0 open-source license. Both the 1-bit binary and 1.58-bit ternary variants are available on Hugging Face, with versions compatible with Apple’s MLX framework for Apple Silicon devices and GGUF format for use with llama.cpp on other hardware. A free developer preview API is also available for testing the model without local download.