Jeffrey Castellano

Find the humanity. Design boundaries. Engineer magic. Rinse. Repeat.

[PrismML @PrismML]

Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone.

Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops.

Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops.
Bonsai 27B changes that.

It comes in two variants:
• Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality.
• 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint.

Everything is open-sourced today under the Apache 2.0 license.

[Ilir Aliu @IlirAliu_]

Forget lidar.
One single camera.
Runs in real time & is open source:

A streaming 3D model that reconstructs scenes live, at ~20 FPS, over long sequences.

End-to-end.

Optimization tricks, cleanup steps?

Nope.

And it beats both streaming and even some offline methods.

Perception is becoming software-first.

Closer to machines that see and understand the world as it unfolds.

Thanks for sharing, @YinghaoXu1

📍Models: https://t.co/gnSDy919eX
Project page: https://t.co/zgpkgBvcik
Code: https://t.co/Js0MzHE387
Paper: https://t.co/FrzMojMMZC

——

Weekly robotics and AI insights.
Subscribe free: https://t.co/9Nm01QUcw3

[Jeffrey Castellano @JeffCastellano]

@felix_trz I'm making a very similar bet right now. Agent first shared underpinnings but using ECS as the mid-speed layer in a three-tier shared memory system. Very cool work.

[Renaud @onirenaud]

Exploring "Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis" running on macOS, no CUDA.

Implemented the training step on Brush (Rust / WGPU / Metal). First test: 1000 iterations/frame.

50 frames → ~20 MB, ~1 ms on GPU.

Plays at 120 fps with three.js.

[Jeffrey Castellano @JeffCastellano]

@xenovacom Gave it a whirl, got real uplift on M1 Max - 95 tok/s average (M1 Max) & 280 TTFT. You've got me beat by about 15 tok/s. 😀 Will have to peek at what you did next week. My 5090, 4070 failed to load. 5080 gave gibberish (weird) sadly on chrome. Metal first or just fighting Dawn?

[Jeffrey Castellano @JeffCastellano]

@mrdoob Nothing to be aware of 😀 - might not have anything worthwhile in it as its webgl with wasm compensating for compute. Just putting it on the radar if it was relevant as I found a couple of clipping gotchas and did a little shadow and experimentation and direct targeting.

[Jeffrey Castellano @JeffCastellano]

@0beqz You in webgpu or stiill webgl these days?

[Jeffrey Castellano @JeffCastellano]

@mrdoob Lots of bugs and mostly abandoned but some good findings too. I was using it mostly inside my SSGI and DDGI work but decided to just plop it separately into a package a year or so ago just to leave it somewhere.

[Jeffrey Castellano @JeffCastellano]

@mrdoob WASM based but maybe some nuggets worth harvest/stealing originally based on the cl2 work. https://t.co/bA9InmYjuP

[Sunag @sea3dformat]

Now we have a point light color node to create the firelight; the shadow also generates transparency according to the smoke.

The example has less than 1k lines of code including HTML.

#threejs #tsl #webgpu

[Jeffrey Castellano @JeffCastellano]

@xenovacom I was doing exactly this too, I did the Gemma 4 kernels a month ago but in one day I got Gemma 12B and diffusion working with WebGPU in my runtime, diffusion got the plug pulled in the final hour last night. They pulled the plug and 4.8 Opus told me it was impossible.

[Jeffrey Castellano @JeffCastellano]

Well, it was fun while it lasted.

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Google Research @GoogleResearch]

Introducing D4RT: A unified AI model for 4D scene reconstruction and tracking across space and time. 🎯 Catch the demo with Skanda Koppula at 12 pm at our #CVPR2026 Google booth kiosk! https://t.co/p6SclNe1zi @GoogleDeepMind

[Jeffrey Castellano @JeffCastellano]

@dangreenheck Congrats! Been watching you grow this out and it really has come a long way. I've been using my own water setup for a long time but I think i have to jump ship for the next use.

[Jeffrey Castellano @JeffCastellano]

@BennJordan gets really creative in making the Unitree backdoors quite visceral. https://t.co/a3IsdoJaPg

[Radiance Fields @RadianceFields]

this is sick

[Zolden @ZoldenGames]

How I created 3D soft body character from simple particle physics.

It took a few months of intense computer science and inventing algorithms to trick a bunch of GPU computed physical particles into dynamically controlling a mesh.

[Jake Wintermute 🧬/acc @SynBio1]

How it feels to do biotech in 2026

[Andrea Tagliasacchi @taiyasaki]

📢📢📢introducing 𝐏𝐨𝐰𝐞𝐫 𝐅𝐨𝐚𝐦

A 3D representation that can be ray traced or rasterized in real time, with NO COMPROMISE in quality.

- Project: https://t.co/LkmVQjkIt2
- arXiv: https://t.co/TtMbyKrvrp

Rasterized at 3DGS-class FPS
Ray traced at Radiant Foam speeds