PrismML's compressed 27B multimodal model — quantized to 1-bit and ternary weights so a model class that used to live in the cloud runs fully on-device, including on a phone.
Last verified · 2026-07-14 · by Moe Ameen
Bonsai 27B is a compressed large language model from PrismML, a US-based AI company that emerged from a team of Caltech researchers with backing from Khosla Ventures, Cerberus, Google, and Samsung. It launched on July 14, 2026, and the headline claim is that it is the first 27B-class model that runs on a phone — a size of model that until now effectively belonged in the cloud or on a workstation.
The trick is aggressive low-bit-width quantization. Bonsai is built on the open Qwen3.6 27B base and shipped in two variants: a ternary version that stores weights as {−1, 0, +1} (about a 5.9 GB footprint) and a 1-bit version that stores weights as binary {−1, +1} (about a 3.9 GB footprint). PrismML reports the ternary variant retains roughly 95% of the full-precision baseline's quality and the 1-bit variant about 90% — so the compression trades a modest amount of accuracy for a very small memory footprint. The 1-bit build is the one that fits within an iPhone 17 Pro's per-app memory budget.
It is multimodal — it accepts images as well as text via a compact 4-bit vision tower — and it carries a long 262K-token context window. Reported speeds range widely by hardware: on a phone the 1-bit variant runs at roughly 11 tokens/sec on an iPhone 17 Pro, while on desktop-class hardware it reaches far higher throughput (PrismML cites up to ~163 tokens/sec on an NVIDIA RTX 5090 and ~87 tokens/sec on an M5 Max for the 1-bit build). Everything is released under the Apache 2.0 license, with the GGUF weights on Hugging Face and a free, limited-time developer-preview API. Because it is a fast-moving first-of-its-kind release, confirm exact specs and any pricing on PrismML's own model card before you build against it.
The novel thing about Bonsai 27B is *where* it runs: in your pocket, offline, on the phone you already film with. That unlocks a specific workflow — capture and draft in the field, then finish and publish back at base. Say you shoot a walkthrough on location with no signal; Bonsai, running locally on the phone, can turn your rough voice notes into three caption angles and a hook while you're still standing there, no data connection needed. What it can't do is become the post. It writes no video, renders no image, sizes nothing for a feed, keeps no consistent brand voice across a week, and publishes nowhere. That's the exact handoff Kompozy is built for. Bring your on-device drafts and your footage into Kompozy, and its Persona Brief rewrites the raw text into your actual brand voice, then fans it into finished formats a phone model can't touch — Clipped Shorts from the footage with burned-in captions, a brand-exact Carousel through HyperFrames, Quote Graphics, Persona and HeyGen avatar video, plus native Text Posts, a Blog Article, and an Email Newsletter.
The pairing plays to Bonsai's strengths — mobility, privacy, and multimodal input. Snap a photo of a whiteboard or a competitor's post and let Bonsai read it and suggest a response on the spot; then Kompozy takes that conclusion and renders it into scheduled, on-brand content across nine social platforms plus email and blog, with Autopilot and a per-post review pipeline. And because Kompozy supports bring-your-own-key on the Founding tier, a team that leans on Bonsai for free, private, on-device drafting keeps that habit while Kompozy owns the media rendering, brand governance, and publishing the little model was never meant to do. Bonsai drafts anywhere, even with the plane in airplane mode; Kompozy turns those drafts into a finished week the moment you're back online.
Bonsai 27B is a compressed large language model from PrismML, launched July 14, 2026. Built on the open Qwen3.6 27B base and quantized to 1-bit or ternary weights, it is described as the first 27B-class model to run on a phone. It is multimodal (reads text and images), has a 262K-token context, and is released under the Apache 2.0 license.
Yes. The 1-bit variant has roughly a 3.9 GB footprint, small enough to fit within an iPhone 17 Pro's per-app memory budget, where PrismML reports around 11 tokens/sec. The ternary variant (about 5.9 GB) needs more headroom but retains higher quality (~95% of the full-precision baseline vs ~90% for 1-bit). On desktop GPUs it runs far faster.
The weights are free under the Apache 2.0 license — you can download the GGUF files from Hugging Face, run them locally, and use the model commercially. PrismML also offered a free, limited-time developer-preview API at launch. Your real cost is the device you run it on.
No. It generates text and reads images on-device, but it renders no video, images, or designs, enforces no brand voice, and publishes to no platform. To turn its drafts into finished, on-brand posts across platforms you pair it with a content engine like Kompozy that renders the media and handles scheduling and publishing.