Bonsai 27B field guide

What the ecosystem is saying about Bonsai 27B

A link first record of The Information's report, PrismML's launch thread, Khosla team amplification, and public reactions to Bonsai 27B.

Published July 15, 2026. Updated when local measurements change.

What is the ecosystem saying about Bonsai 27B?

The main public story has two parts. The Information reported that Apple had talked with PrismML about possible uses of its compression work. PrismML then released the 27B checkpoints, benchmarks, and demos. Public reactions focused on the phone size claim, local agents, and the need to test quality outside vendor benchmarks.

The Information is the centerpiece

Aaron Tilley reported on July 9 that PrismML had run a compressed Qwen3.6 27B model on an iPhone 17 Pro and that Apple had discussed possible uses of the technology with the company. The article is behind a subscription wall, so this page links to it and does not copy its text.

PrismML released Bonsai 27B on July 14. Its own announcement gives the model sizes, benchmark claims, formats, and license. These are company claims until independent tests reproduce them.

The Information report by Aaron Tilley July 9, 2026
PrismML launch announcement July 14, 2026
CNBC follow-up by MacKenzie Sigalos July 14, 2026

PrismML's public launch thread

PrismML used six posts to separate the launch, benchmark, phone memory, agent demo, intelligence density, and local agent arguments. The links below go to the original posts.

What the Khosla team amplified

Kanu Gulati is a Khosla Ventures partner and lists PrismML among the companies he led. His public feed reshared the reactions from Ion Stoica, Together AI, NVIDIA RTX Spark, and other launch participants. That shows investor support and distribution. It is not an independent benchmark.

Independent reactions and open questions

Vaibhav Sisinty summarized the size and benchmark claims for a broad builder audience. Other early posts focused on whether the release changes local agent economics. The strongest open question is still task level quality, especially for tool use and long agent loops.

Our local work answers a narrower question. The public ternary MLX checkpoint loads on a 24 GB M4 Pro Mac, generates at about 21 to 23 tokens per second in controlled 4K runs, and can return a structured tool call. It does not verify the iPhone result or the full vendor benchmark suite.

Questions people ask

Did Apple confirm a deal with PrismML?

No public deal has been confirmed. Reports describe early evaluation and discussions.

Did Khosla Ventures publish an independent benchmark?

No public independent benchmark was found. The public record shows investment support and amplification of the launch.

Are all X reactions independent?

No. PrismML's posts are company statements, and investor or platform partner posts have their own interests. This page labels each group.

Primary sources

Vendor claims on this page are labeled as PrismML claims. Local results are added only after a saved trace completes.