A 27-billion-parameter language model, the kind that used to need a rack of GPUs, is now reportedly running entirely inside a phone in your pocket. According to coverage from Wccftech and Notebookcheck published this week, Apple’s upcoming A20 Pro chip has been shown running a 27B-parameter model at roughly double the token-generation speed of last year’s A19 Pro, the chip inside the iPhone 17 Pro. Apple itself has not issued an official statement confirming the specific figures behind the claim. The numbers come from independent developers and a small AI compression startup, not from Cupertino’s marketing department, which is precisely why the story has spread so fast among engineers who build on-device AI for a living.
The chip at the center of the story, the A20 Pro, is described in current reporting as Apple’s first 2-nanometer iPhone chip, expected to ship inside the iPhone 18 Pro generation. Before that hardware landed in more hands, an iPhone enthusiast named Adrien Grondin reportedly ran a 27-billion-parameter model called Bonsai 27B entirely locally on an iPhone 18 Pro, no cloud calls, no API keys, just the phone’s own silicon doing the inference work. That demonstration followed months of groundwork from a small AI shop called PrismML, which had already gotten a related 27B model running on last year’s iPhone 17 Pro. Put together, the two data points tell a story about how fast on-device AI is moving, and how much of that progress is happening outside Apple’s own press releases.
What Wccftech and Notebookcheck Are Actually Reporting
Strip away the headline and the underlying claim is narrower than it sounds, but no less notable. Wccftech and Notebookcheck are reporting on a chain of independent demonstrations, not an Apple keynote. The chain starts with Alibaba’s Qwen team, which released Qwen3.6-27B, a dense 27-billion-parameter model, as an open-weight download. PrismML, an AI development company, then took that roughly 54GB model and compressed it to under 4GB while keeping all 27 billion parameters active at once, a process the company demonstrated running on an iPhone 17 Pro alone. A related, smaller-footprint version of that compressed model, called Bonsai 27B, needs about 3.9GB of memory, small enough to fit inside what an app is normally allowed to use on an iPhone 17 Pro.
The A20 Pro piece of the story is the newer, less-documented half. Grondin’s report of running Bonsai 27B locally on an iPhone 18 Pro, the phone expected to carry the A20 Pro chip, is what pushed outlets to frame this as an A20 Pro story rather than just another PrismML update. Neither Wccftech nor Notebookcheck has published a fully independent, apples-to-apples tokens-per-second benchmark comparing A19 Pro and A20 Pro on the identical model and prompt set. Readers should treat the “double the speed” framing as the outlets’ interpretation of early, developer-run tests, not as a number Apple has published or verified.
Inside the Numbers: How Bonsai 27B Actually Performs
The most concrete numbers in this story come from the A19 Pro generation, where multiple independent parties have now measured Bonsai 27B under real conditions. Kingy AI, an analytics outlet tracking on-device AI performance, reports that independent developers clocked Bonsai 27B at roughly 8 to 9 tokens per second on 8GB iPhones, which covers the base iPhone 17 Pro. Separately, RunAnywhere, a company that benchmarks local AI runtimes, measured about 10 tokens per second running the same model on an iPhone 17 Pro. PrismML’s own measurement, using Apple’s MLX Swift framework, came in at 11.0 tokens per second on an iPhone 17 Pro Max.
Those numbers cluster tightly, which is a good sign for anyone skeptical of cherry-picked demo footage. Three separate parties, using slightly different hardware and tooling, all landed in the same 8-to-11 tokens-per-second band on A19 Pro silicon. For context, a comfortable reading pace for most people is somewhere around 3 to 5 words per second, so even the low end of that range is fast enough to feel conversational rather than sluggish, though it is well short of what a cloud-hosted model serves back over an API.
| Device / Chip | Tester | Tokens per second | Model configuration |
|---|---|---|---|
| iPhone 17 Pro (A19 Pro, 8GB) | Independent developers, via Kingy AI | 8-9 | Bonsai 27B, compressed |
| iPhone 17 Pro (A19 Pro) | RunAnywhere | 10 | Bonsai 27B |
| iPhone 17 Pro Max (A19 Pro) | PrismML, via MLX Swift | 11.0 | Bonsai 27B |
| iPhone 18 Pro (A20 Pro) | Adrien Grondin (independent) | Not publicly benchmarked yet | Bonsai 27B, full local run confirmed |
| NVIDIA GeForce RTX 5090 (desktop) | Reported via coverage of Bonsai 27B | 163 (1-bit variant) / 134 (3-level variant) | Bonsai 27B, quantized variants |
Notice the gap in that A20 Pro row. That gap is the whole story, and it is also the biggest reason to read the “2x faster” headline with a healthy amount of caution. Grondin’s demonstration confirms the A20 Pro can carry the model’s full 27 billion parameters and run inference without crashing or throttling into uselessness. It does not, on its own, hand us a tokens-per-second figure that outlets can responsibly publish as fact. The desktop GPU row is there for scale, not comparison. A RTX 5090 pulling well over 100 tokens per second on quantized Bonsai variants is a reminder of just how much power a phone chip is giving up to fit in a pocket, even while closing the gap faster than most people expected.
How PrismML Shrank a 54GB Model Down to Under 4GB
The compression work is arguably the more technically interesting part of this story, since chip speed only matters if the model can physically fit in memory in the first place. Qwen3.6-27B ships at around 54GB in its original form, a size that rules out virtually every consumer phone on the market today, iPhone included. PrismML’s contribution was getting that model down to less than 4GB while keeping all 27 billion parameters active simultaneously, rather than swapping parameters in and out of memory or dropping a chunk of the network entirely.
That is roughly a 13-to-1 size reduction. Aggressive quantization, the technique of storing each parameter using far fewer bits than the 16 or 32 bits a model is normally trained with, is the standard tool for this kind of shrinkage, and it is almost certainly the tool PrismML leaned on given the naming of Bonsai 27B’s “1-bit” and “3-level” variants referenced in coverage of its desktop GPU performance. The tradeoff with quantization is always the same: push the compression too far and the model starts making more mistakes, repeating itself, or losing coherence on longer responses. Coverage of Bonsai 27B does not report an accuracy or benchmark-score drop-off figure for the compressed version compared to the original Qwen3.6-27B, so it stays an open question exactly how much quality got traded for that 13-to-1 size cut.
Why 27 Billion Parameters Is the Number Everyone Is Watching
Most on-device AI features shipped by phone makers up to this point have leaned on far smaller models, often in the single-digit-billion-parameter range, because that is what fits comfortably in a phone’s memory and thermal budget. A 27B model landing on-device, even a compressed one, is a jump of an order of magnitude past what shipped in most 2025-era on-device AI features. That is why an independent demo, run by one enthusiast on one phone, was enough to get picked up by Wccftech and Notebookcheck rather than staying a niche AI-Twitter story.
A19 Pro vs A20 Pro: What’s Actually Confirmed
It is worth separating the chip facts that are solid from the ones still riding on rumor and demo footage. The A19 Pro is confirmed as the chip inside the iPhone 17 Pro, currently on sale. The A20 Pro is described in current reporting as Apple’s first 2-nanometer iPhone chip, tied to the iPhone 18 Pro generation. Beyond the process node and the model-name pairing, Apple has not published core counts, neural engine specs, or official AI throughput numbers for the A20 Pro, so any specific performance multiplier attached to it right now is coming from outside reporting, not from Apple’s own spec sheet.
| Chip | iPhone model | Process node | 27B model status |
|---|---|---|---|
| A19 Pro | iPhone 17 Pro / Pro Max | Not specified in confirmed sources | Confirmed running Bonsai 27B, 8-11 tokens/sec across three independent testers |
| A20 Pro | iPhone 18 Pro | First 2nm iPhone chip, per current reporting | Confirmed running Bonsai 27B locally (Grondin); speed advantage reported by Wccftech and Notebookcheck, not yet independently benchmarked |
That table looks thin on the right side for a reason. This is a fast-moving story where the hardware claim (2nm process, a real speed jump) is more solid than the specific performance multiplier being quoted around it. Readers chasing an exact, Apple-verified benchmark number for the A20 Pro running Bonsai 27B or Qwen3.6-27B will not find one yet. What they will find is a working local demo, on real hardware, which is itself a meaningful checkpoint even without a precise speed figure attached.
Why On-Device 27B Models Matter for Privacy and Security
For a security and privacy audience, the interesting part of this story is not really the tokens-per-second race. It is what running a model this large locally removes from the equation: the network round trip. A cloud-hosted AI assistant sends your prompt, and whatever context it’s carrying, to a remote server, where it sits in a request log, potentially gets used for further training, and adds a network path an attacker could theoretically intercept, phish, or subpoena. A model that runs entirely inside the phone’s own memory and compute never has to make that trip at all.
That matters more the bigger and more capable the local model gets. A 3B-parameter on-device assistant is useful for autocomplete and simple summarization, but a 27B-parameter model starts to approach the kind of reasoning and context-handling that people currently reach for cloud assistants to get. If Apple, or any phone maker, can push genuinely capable models onto the device itself, it changes the privacy calculus for an entire category of AI features, from health-adjacent journaling apps to on-device translation of sensitive documents, without a single byte of that content needing to leave the handset.
The flip side deserves a mention too. A phone holding a 27B model locally is also a phone holding a much bigger, more valuable target if that device is ever lost, stolen, or physically compromised. The security calculation shifts from “is my data safe in transit and on someone else’s server” to “is my device’s local storage and memory protection strong enough,” which puts even more weight on Apple’s existing Secure Enclave and on-device encryption work.
Market Impact: Apple’s On-Device AI Bet Versus the Field
Apple has spent the last two years being characterized, fairly or not, as a step behind on generative AI compared to Google and its Gemini rollout or Samsung’s Galaxy AI push. A story like this one, even if it originated from independent developers rather than an Apple keynote, works in Apple’s favor because it reframes the conversation around raw silicon capability rather than app-store feature parity. If a 2nm A20 Pro really can run a 27B model at meaningfully higher speed than last year’s flagship chip, that is a hardware moat few competitors can match quickly, since it depends on chip design lead time measured in years, not a software update.
Competing phone makers generally lean on smaller, purpose-built on-device models rather than chasing raw parameter count, favoring efficiency and battery life over headline-grabbing size. That is a reasonable strategy when the on-device use case is autocomplete, photo tagging, or simple voice commands. It becomes a harder sell if consumers start associating “runs a real 27B model locally” with Apple hardware specifically, the way “best camera” or “best battery life” claims have shaped past phone-buying decisions. Expect rival chipmakers to respond with their own on-device benchmark claims well before their next flagship silicon actually ships.
Historical Context: From Siri’s Cloud Dependence to Local LLMs
It is easy to forget how recently “on-device AI” meant almost nothing beyond wake-word detection and basic autocorrect. Siri, for most of its history, has leaned heavily on cloud processing for anything beyond trivial commands, a design choice that made sense when phone chips simply could not run a useful language model locally. Apple’s own on-device push accelerated with the Apple Intelligence rollout, which introduced smaller foundation models running locally for tasks like writing tools and notification summaries, while routing more demanding requests to Apple’s Private Cloud Compute infrastructure.
The Bonsai 27B and Qwen3.6-27B story represents a different lineage entirely: independent researchers and a small compression startup pushing far past what any phone maker has shipped as an official feature, using open-weight models Apple did not build and does not control. That is a meaningful shift in how phone AI capability gets discovered and reported. The most notable local-AI news of the week did not come from Apple’s newsroom. It came from a developer’s personal demo and a startup’s compression pipeline, running on hardware Apple designed for entirely different, officially unannounced purposes.
Independent Testing: RunAnywhere, Kingy AI, and Artificial Analysis
Three named organizations show up repeatedly across the reporting on this story, and each brings a slightly different angle. RunAnywhere focuses on measuring how local AI runtimes actually perform on consumer hardware rather than in a lab, which is why its iPhone 17 Pro figure lines up closely with numbers reported by independent developers rather than a single vendor’s marketing benchmark. Kingy AI aggregates and reports on those independent developer tests, giving a wider, though less centrally controlled, view of real-world Bonsai 27B performance across different iPhone configurations.
Artificial Analysis, an outlet that has published measurements of AI performance on iPhones more broadly, adds a layer of third-party scrutiny to a space that could otherwise run entirely on self-reported vendor numbers. None of these three groups is Apple, and none of them is PrismML, which matters: a claim that only comes from the company that built the compressed model carries less weight than one echoed by outside testers using their own hardware and their own stopwatch. The rough agreement across RunAnywhere, Kingy AI’s aggregated developer data, and PrismML’s own MLX Swift measurement is what turns this from a marketing claim into something closer to a verified, if still narrow, result.
Running a Quantized Model Locally: A Simplified Example
For engineers curious about the mechanics, loading a quantized model through Apple’s MLX Swift framework, the same tooling PrismML used for its iPhone 17 Pro Max measurement, generally follows a pattern like the simplified example below. This is illustrative of the general MLX Swift workflow, not a reproduction of PrismML’s proprietary compression pipeline.
import MLX
import MLXLLM
// Load a locally-stored, pre-quantized model bundle
let modelPath = "Bundle/bonsai-27b-quantized"
let model = try await LLMModelFactory.shared.loadModel(from: modelPath)
// Run inference entirely on-device, no network call
let prompt = "Summarize this document in three sentences."
let result = try await model.generate(prompt: prompt, maxTokens: 200)
print(result.text)
print("Tokens/sec: \(result.tokensPerSecond)")The actual work of getting a 54GB model down under 4GB happens well before this stage, in the offline quantization and compression pipeline, not in the loading code itself. Once a model is compressed and bundled this way, the runtime code that loads and queries it stays comparatively simple, which is part of why independent developers have been able to reproduce these on-device 27B demonstrations without needing Apple’s direct involvement.
Risks and Unanswered Questions
A few gaps in this story are worth flagging plainly rather than glossing over. First, there is no independently published accuracy benchmark comparing compressed Bonsai 27B against the original, uncompressed Qwen3.6-27B, so it is not yet clear how much reasoning quality was traded away for that roughly 13-to-1 size reduction. Second, none of the reporting includes battery drain or thermal throttling data for sustained local inference, and a phone running a 27B model for an extended session is a very different thermal situation than a quick 30-second demo clip.
Third, and most relevant to the headline itself, there is still no independently verified, apples-to-apples benchmark putting A19 Pro and A20 Pro side by side on the same model, same prompt, and same conditions. Until that test exists and gets published by a party with no stake in the outcome, the “double the speed” framing should be read as an early, directionally plausible claim rather than a settled fact.
What This Means for Developers Building On-Device AI
For app developers, the practical takeaway is less about the exact tokens-per-second figure and more about a shifting baseline. If a 27B-class model can realistically run on current-generation iPhone hardware at a usable speed, the calculus for building “local-first” AI features, ones that work offline, keep user data on-device, and don’t rack up inference API bills, starts to look far more attractive than it did even a year ago. Expect more indie developers to experiment with MLX-based local inference over the next few months, using exactly the kind of compression techniques PrismML has been demonstrating publicly.
It also raises the bar for what “on-device AI” means as a marketing claim. A feature that quietly ships a 1-billion-parameter classifier is a very different product than one running a compressed 27-billion-parameter general-purpose model, even if both get labeled “runs locally on your phone” in an app store listing. Reviewers and security researchers alike will likely start asking for parameter counts and memory footprints the same way they now ask about encryption standards.
Predictions: Where On-Device AI Goes From Here
- Expect an independent, third-party benchmark directly comparing A19 Pro and A20 Pro on an identical local model within the next few months, likely from one of the same groups already active in this space, such as Artificial Analysis or RunAnywhere.
- PrismML and similar compression outfits will likely push toward even larger models, testing whether 30B-plus parameter counts can be squeezed into phone-friendly memory footprints using more aggressive quantization.
- Rival chipmakers supplying Android flagships will respond with their own on-device AI performance claims well ahead of their next silicon generation actually shipping, to avoid ceding the “best local AI phone” narrative to Apple.
- Accuracy and quality benchmarks for these heavily compressed models will become the next battleground, since raw tokens-per-second numbers only tell part of the story if the compressed model’s answers are noticeably worse.
- Apple will likely address on-device model capability more directly in future Apple Intelligence updates, rather than leaving the narrative entirely to third-party developer demos, given how much attention this story has already drawn.
Frequently Asked Questions
Has Apple officially confirmed the A20 Pro runs a 27B model at double the speed of the A19 Pro?
No. Apple has not issued an official statement confirming these specific performance figures. The claims trace back to third-party reporting from Wccftech and Notebookcheck, covering independent developer demonstrations, not an Apple benchmark disclosure.
What is Bonsai 27B?
Bonsai 27B is a 27-billion-parameter AI model, described as a smaller version of Alibaba Qwen’s Qwen3.6-27B, built by PrismML to run locally on phone hardware with roughly 3.9GB of memory usage.
What chip is inside the iPhone 17 Pro?
The iPhone 17 Pro uses Apple’s A19 Pro chip, which is the same chip that independent testers used to measure Bonsai 27B running at roughly 8 to 11 tokens per second.
Is the A20 Pro Apple’s first chip built on a 2-nanometer process?
Current reporting describes the A20 Pro as Apple’s first 2-nanometer iPhone chip, expected in the iPhone 18 Pro generation, though Apple has not published full technical specifications for it yet.
How much did PrismML compress the original Qwen3.6-27B model?
PrismML reports compressing Qwen3.6-27B from roughly 54GB down to under 4GB while keeping all 27 billion parameters active simultaneously, a reduction of roughly 13 to 1.
Can I run a 27B-parameter model on my own iPhone today?
Independent developers have reportedly run compressed versions like Bonsai 27B on iPhone 17 Pro and iPhone 18 Pro hardware using frameworks such as MLX Swift, but this is not a shipped, officially supported Apple feature, and results may vary based on device memory and configuration.
Why does running a large AI model locally matter for privacy?
A model that runs entirely on-device never has to send prompts or context to a remote server, which removes the network transmission and server-side logging risks that come with cloud-hosted AI assistants.
How fast is 8 to 11 tokens per second in practical terms?
That range is roughly in line with or faster than typical human reading speed, meaning generated text should appear at a comfortable, conversational pace rather than feeling sluggish, though it remains slower than most cloud-hosted API responses.
