Bonsai 2 Bonsai 2 is a family of compact ternary AI models in 2B, 4B, and 9B sizes based on Qwen3.5. Optimized for local on-device inference, they deliver reasoning, visual understanding, and tool calling in footprints as small as 0.43GB, with up to 40% fewer tokens on everyday tasks and high throughput on phones and PCs.

Visit product

Featured media

Cosigns Received

Sudo su
“i had to check the file size twice prismml dropped bonsai 2 at 2b, 4b and 9b today, ternary, built on qwen 3.5, and the 9b is a 2gb file, the qwen 3.5 9b it came from is 17.9gb the acceleration in open source right now is insane, bonsai 2 27b came out a month ago at 5.95gb and it's already at 4.8m downloads, i run it on a 3060 12gb at 50.1 tok/s with my kernel, this week alone i counted 11 releases from a 5M forecaster to a 600b moe, and now a 9b with vision and tool calls fits on a phone the 9b goes through my tests this week, numbers from my own cards soon”
Source
Chirag
“PrismML's Ternary Bonsai 2 2B and 4B now run in a browser tab, in LocalMind: 430 MB and 950 MB, tool calls included, and nothing leaves your device. Token for token the same as PrismML's own llama.cpp on our test prompts. https://localmind.naklitechie.com”
Source
Som Dutt | AI Analyst
“The 2B version has an even bigger advantage over the 2B version of Qwen3.5. It uses less memory (0.43 GB vs 3.76 GB) and has better accuracy on mathematical problems (81.57 vs 75.68) and instruction following (66.89 vs 55.06).”
Source
atomic.chat
“Run Ternary Bonsai 2 locally in Atomic Chat 🪴 @PrismML shrank Qwen3.5 7x so the 9B goes from 17.9 GB to just 2.42 GB and it keeps up or outperforms the original model on benchmarks Run AI models locally -> https://atomic.chat”
Source
Doğukan
“PrismML's Bonsai 2 small models are out -- 2B, 4B and 9B, real ternary, GGUF + MLX. (configs below for all three) a 9B in 2 GB, a 2B in 430 MB. with 32K context the 2B runs in under 1 GB of VRAM. that's on-device territory -- laptops, old cards, macs, fully offline. 9B on my RTX 3060 with PTQ1_0: ~114 tok/s decode, ~1,800 tok/s prefill, full 262K context + vision in ~7.5 GB. PTQ1_0 is the smallest file and on my 3060 it's also the fastest, ~15-20% faster decode than PQ2_0. needs PrismML's llama.cpp fork, stock llama.cpp won't load it. day one config, expect it to move. config: llama-server -m Ternary-Bonsai-2-9B-PTQ1_0.gguf --mmproj Ternary-Bonsai-2-9B-mmproj-Q8_0.gguf -ngl 999 -c 262144 -fa on --jinja -np 1 --cache-type-k q8_0 --cache-type-v q8_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 1.0 --repeat-penalty 1.0 --reasoning on”
Source
Romain Hedouin 🇫🇷 Unit Police
“Just tried Bonsai 2 9B on a Macbook Air 16 GB (the cheapest Macbook) For the first time ever, I'm blown away by a small LLM”
Source