PrismML {bot}

Today, we’re releasing Bonsai 2 2B, 4B, and 9B. Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices: • 2B: 0.43GB • 4B: 0.95GB • 9B: 2.0GB (1/5)

The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware. Available today under Apache 2.0. (5/5)

We’re also launching an optimization competition with @darkbloomai to make Bonsai 2 9B faster. Check it out on @YukonResearch! https://twitter.com/PrismML/status/2110854487942578445 Video: https://twitter.com/prismml/status/2110854487942578445

https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-small-whitepaper.pdf Discord: https://discord.com/invite/prismml Join us: https://prismml.com/careers (4/5)

Alongside the release, we're launching an open challenge on @YukonResearch to make Bonsai 2 9B even faster. You can also try it today on @DarkbloomAI. (5/5)

Blog: https://prismml.com/news/bonsai-2-family Models: https://huggingface.co/collections/prism-ml/bonsai-2-small Darkbloom API: https://console.darkbloom.dev/models Demo: https://prismml-eng.github.io/Talk/ Try on iPhone: (2/5)

https://apps.apple.com/us/app/locally-ai-by-lm-studio/id6741426692 Instructions and Examples: https://github.com/PrismML-Eng/Bonsai-demo/ Docs: https://docs.prismml.com/get-started/introduction.md Whitepaper: (3/5)

Building with Bonsai? We work with teams on domain-specific post-training and hardware optimization to meet memory, latency, and power constraints. Get in touch: contact@prismml.com.

prismml.comPrismML — CareersExplore open AI/ML, platform, product, developer relations, and recruiting roles at PrismML in Pasadena or San Francisco.

Bonsai 2 9B, 4B, and 2B are available under Apache 2.0. We look forward to seeing what you build. (1/5)

In this demo, Bonsai 2 processes a confidential data room on-device, replacing sensitive details with codes before a cloud model analyzes the material and drafts a due diligence memo. (3/4)

It keeps sensitive information on-device and sends just 8.6% of the data to the cloud, reducing cloud costs by 86%. Video: https://twitter.com/prismml/status/2110823968831832189 (4/4)

Here is Bonsai 2 9B as part of a hybrid AI workflow on Mac. (1/4)

Bonsai 2 can serve as the local intelligence layer in systems that combine on-device and cloud inference. This can support privacy-focused applications by giving developers more control over where information is processed and what leaves the device. (2/4)

Here is Bonsai 2 9B running locally on an iPhone, performing agentic tasks. It’s a practical example of bringing reasoning and tool use directly into a mobile application. (1/2)

On iPhone 18 Pro Max, Bonsai 2 9B reaches up to 34 toks/sec, while the 2B model reaches up to 120 toks/sec. https://twitter.com/PrismML/status/2110823966717927773 Video: https://twitter.com/prismml/status/2110823966717927773 (2/2)

The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait. (3/5)

These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables. (4/5)

For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. (2/5)

We have been very excited to see how the community has been utilizing our Bonsai 2 model, well beyond use cases that we had originally envisioned it for. We’ve spent the last few days putting Bonsai 2 27B through various examples inspired by what the community has shown us. (1/3)

Code: https://github.com/PrismML-Eng/Bonsai-demo/tree/agent-demo-bonsaios https://twitter.com/PrismML/status/2103259930614767643 Video: https://twitter.com/prismml/status/2103259930614767643 (2/2)

We’ll keep sharing both what the model does well and where we’re working to make it better. More updates soon.

Bonsai scored in the upper end of the human bronze-medal range and retained 95% of Qwen’s IMO score, while completing the problems in 70% of the time. Code: https://github.com/PrismML-Eng/Bonsai-demo/tree/imo2026-demo Video: https://twitter.com/prismml/status/2103259928274317349 (2/2)

github.comGitHub - PrismML-Eng/Bonsai-demo at agent-demo-bonsaiosBonsai Demo. Contribute to PrismML-Eng/Bonsai-demo development by creating an account on GitHub.

Starting from an empty workspace, Bonsai 2 27B built a browser-based desktop in a single HTML file, then iteratively fixed issues across multiple rounds of feedback, from broken windows and UI behavior to complete functionality and final styling. (1/2)

Code: https://github.com/PrismML-Eng/Bonsai-demo/tree/agent-demo-skateboard Video: https://twitter.com/prismml/status/2103259925925564918 (3/3)

github.comGitHub - PrismML-Eng/Bonsai-demo at imo2026-demoBonsai Demo. Contribute to PrismML-Eng/Bonsai-demo development by creating an account on GitHub.

We also tested Ternary Bonsai 2 27B on the 2026 IMO problems against full-precision Qwen3.8 27B (54GB) and Gemma 4 12B QAT (~7GB). No internet. No tools. 131K-token reasoning budget. (1/2)

github.comGitHub - PrismML-Eng/Bonsai-demo at agent-demo-skateboardBonsai Demo. Contribute to PrismML-Eng/Bonsai-demo development by creating an account on GitHub.

First up: building a 3D skateboard game from scratch in a single HTML file. The first attempt isn't perfect. Instead of restarting a better generation, we keep the same conversation going, point out the issues, and give the model feedback. (1/3)

It progressively fixes the gameplay, visuals, and controls over several iterations. That iterative loop being able to take feedback, debug its own work, and improve an existing artifact is much closer to how we think language models are most useful in practice. (2/3)

These examples showcase the strengths, and some of the shortcomings–particularly on longer multi-turn agentic workflows, and will be extremely helpful as we continue to improve. (2/3)

We’ll be sharing a few of those demos, along with some practical guidance on the settings and prompting that get the best results from the model. Demo repo: https://github.com/PrismML-Eng/Bonsai-demo/tree/main (3/3)

techcrunch.comPrismML brings its tiny LLMs to Qualcomm-powered smart glasses | TechCrunchPrism's larger goal is open-weight AI that runs on devices and makes better use of the computing power they already have.

Great to see @TechCrunch cover our work with @Qualcomm! Read here: https://techcrunch.com/2026/09/24/prismml-brings-its-tiny-llms-to-qualcomm-powered-smart-glasses/

Today at Snapdragon Summit, we’re announcing 1-bit Bonsai vision-language model running locally on AI smart glasses powered by Snapdragon® AR1 Gen 1 Platform. (1/3)

prismml.comPrismML — PrismML Brings 1-Bit Bonsai Models to AI Smart Glasses Powered by Snapdragon1-bit Bonsai dramatically increases the on-device intelligence of smart glasses, fitting 4× larger models onto the same memory footprint

Read the full announcement: https://prismml.com/news/prismml-brings-1-bit-bonsai-models-to-ai-smart-glasses-powered-by-snapdragon

Proud to be working with @Qualcomm to bring increasingly capable AI directly onto highly constrained devices.

Today at Snapdragon Summit, we’re announcing 1-bit Bonsai vision-language model running locally on AI smart glasses powered by Snapdragon® AR1 Gen 1 Platform. (1/3)

With Qualcomm Technologies, we optimized Bonsai for Snapdragon Hexagon™ NPU, delivering 4x lower memory usage and 2x faster token generation for equivalent intelligence, making it possible to bring significantly more capable AI to run directly on highly constrained wearable devices. (2/3)

We believe model-hardware co-design will be critical to bringing increasingly capable AI from the cloud onto the devices people use every day. (3/3)

Today, we’re announcing Ternary Bonsai 2 27B. Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance. (1/3)

prismml.comPrismML — CareersExplore open AI/ML, platform, product, developer relations, and recruiting roles at PrismML in Pasadena or San Francisco.

Links: Blog: https://prismml.com/news/bonsai-2-27b Whitepaper: https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-2-27b-whitepaper.pdf Models: https://huggingface.co/collections/prism-ml/bonsai-2 WebGPU Demo: (1/2)

https://huggingface.co/spaces/webml-community/ternary-bonsai-2-webgpu-kernels GitHub: https://github.com/PrismML-Eng/Bonsai-demo/ Docs: https://docs.prismml.com Discord: https://discord.gg/prismml Join us: https://prismml.com/careers (2/2)

Here is Bonsai 2 27B running a computer-use workflow locally on the NVIDIA GeForce RTX 5090 GPU. Video: https://twitter.com/prismml/status/2100692255057367487 (2/2)

Building with Bonsai? We work with teams on domain-specific post-training and hardware optimization to meet memory, latency, and power constraints. Get in touch: contact@prismml.com.

Here is Ternary Bonsai 2 27B running an agentic coding workflow with Cline on an NVIDIA GeForce RTX 5090 GPU. Video: https://twitter.com/prismml/status/2100692253090189324 (2/2)

Computer use is another strong test. The model has to repeatedly interpret state, choose the next action, and stay coherent across a long sequence of steps, exactly where small capability gaps become visible. (1/2)

Many low-bit models become deployable only with a meaningful quality tradeoff. Bonsai 2 pushes the frontier toward significantly higher capability at the same memory footprint. (2/2)

This is where higher retention matters most: fewer derailments across multi-step tasks, fewer silent failures as the state evolves, and better consistency from one decision to the next. (1/2)

Ternary Bonsai 2 27B is available today under Apache 2.0. (3/3)

On an intelligence-density basis, Bonsai 2 27B is a clear outlier relative to both full-precision models and other low-bit alternatives. (1/2)