Bonsai 27B Claims a 3.9 GB Footprint and 11 Tokens Per Second on an iPhone 17 Pro-class Device

#ai #iphone #machinelearning #llm

Bonsai 27B can plausibly run on a phone under PrismML’s stated conditions, because PrismML says its 1-bit MLX release is about 3.9 GB and reaches about 11 tokens per second on an iPhone 17 Pro Max. The important catch is simple: that phone result is PrismML’s own launch claim on a high-end iPhone-class device, not an independent long-session mobile benchmark.

PrismML is a model compression startup pitching Bonsai as a way to squeeze large models into much smaller runtimes. On July 14, 2026, it said its new Bonsai 27B line is derived from Qwen3.6-27B and turns a 27.8 billion-parameter multimodal model into a phone-scale release without fully falling apart on benchmarks.

PrismML’s own numbers make the compression look dramatic. The public MLX 1-bit model card lists a 3.9 GB download, while PrismML’s unpacked repository shows the full model at 54.7 GB. That is roughly a 14x size reduction from the unpacked release to the 1-bit mobile-oriented version. For a 27B-class model, that is the whole story: it moves from obviously not phone-friendly to at least physically loadable on top-tier phones.

Bonsai 27B’s claimed phone-scale footprint and speed

PrismML’s launch post says Bonsai 27B was built for Apple’s MLX stack, and the 1-bit model card says it runs at around 11 tokens per second on iPhone 17 Pro Max. An independent 9to5Mac report surfaced the same framing: PrismML is not saying “phones” in general, but a current high-end iPhone with the right memory headroom.

That memory headroom matters more than the headline does. A 3.9 GB weight file is not the whole runtime budget; KV cache and activations still consume memory during inference. That is the practical point behind our earlier look at the Bonsai 27B on iPhone memory and speed claim: fitting the weights is necessary, not sufficient.

A short comparison makes the claim clearer:

Model artifact	Reported size / speed
Bonsai 27B MLX 1-bit	3.9 GB, ~11 tok/s on iPhone 17 Pro Max
Bonsai 27B unpacked full model	54.7 GB

What is missing is just as important. No primary source here provides sustained battery draw, thermal throttling, or long-session latency data for phone use. A flashy on-stage or launch-demo throughput number is not yet the same thing as “comfortable daily local AI on a phone.”

What 27B-class performance means in PrismML’s benchmark table

PrismML’s quality claim rests on a benchmark-retention argument, not on parity. In its launch announcement, the company says the 1-bit Bonsai 27B keeps about 89.5% of the FP16 model’s average score across 15 benchmarks. That is strong for this level of compression. It is also a very specific claim: retained average score, across PrismML’s chosen set, under PrismML’s evaluation setup.

The 1-bit model card publishes the benchmark table PrismML is leaning on. The headline is not that Bonsai beats full-precision Qwen3.6-27B. It does not. The headline is that a heavily compressed derivative still tracks surprisingly close to the original across a broad test set.

PrismML also points to a broader Bonsai family that includes related variants in its Hugging Face collection, including ternary releases. The launch materials say the ternary version preserves more quality than the strict 1-bit release, which is exactly what you would expect: more representational room, less brutal compression. But the most detailed ternary quality claims in this source set come from PrismML’s own published model materials rather than third-party testing.

That makes the right reading fairly plain. “27B-class performance” here means “benchmark retention close enough to remain recognizably in the original class,” not “full 27B performance at full precision.” If you care about local coding or serious assistant use, that distinction is not nitpicking; it is the difference between an intriguing edge deployment and a drop-in replacement for a normal 27B setup. Readers weighing that tradeoff should also look at a broader best local coding model guide rather than assuming one compression trick settles the category.

The practical limits of “runs on a phone”

The clean answer is yes, with conditions. Bonsai 27B appears to really be a released model family with a public 1-bit MLX card, not just a teaser, and PrismML has published enough artifacts to make the claim concrete: the July 14 release post, the 1-bit MLX repository, the unpacked model repository, and the family collection page.

But “runs on a phone” still means something narrower than the phrase suggests. PrismML’s public materials tie the demo to an iPhone 17 Pro-class target running MLX, not to mainstream Android devices, older iPhones, or phones in general. And the only speed figure in this source set — about 11 tokens per second — comes from PrismML’s own launch materials.

That leaves three practical unknowns unresolved. First, sustained performance: phones heat up, and thermal throttling can turn a neat demo into a sluggish one. Second, usable context length under mobile memory limits: a small weight file does not stop KV cache growth as chats get longer. Third, battery cost: no source here says what repeated local inference does to an actual day’s charge.

So the right verdict is narrower than the hype and stronger than the skepticism. Bonsai 27B probably does clear the bar of “can a compressed 27B-derived model run locally on a top-end phone,” but PrismML has not yet shown enough independent evidence to upgrade that into “27B models are now practically mobile.” The next milestone is obvious: third-party testing on real devices, over long sessions, with thermals and battery measured instead of implied.

Key Takeaways

PrismML announced Bonsai 27B on July 14, 2026, and the release appears to be real rather than a teaser because public model artifacts are already live.
The 1-bit MLX Bonsai 27B model card lists a 3.9 GB size and about 11 tokens per second on an iPhone 17 Pro Max.
The unpacked full model repository shows 54.7 GB, making the compressed 1-bit release roughly 14 times smaller.
PrismML says the 1-bit model retains about 89.5% of the FP16 parent model’s average score across 15 benchmarks, which supports the “27B-class” framing without implying full parity.
No source here provides independent long-session phone data for battery, thermals, or sustained latency, so the mobile claim is still partly a controlled-demo claim.