@ZAlvin39105 Great work Alvin, some of my work involves looped models so I'll be taking a good look later! Abstract already caught my eye. test-time adaptation in recurrent transformers... very, very interesting.
Cool if I DM?
English
David Kogan
233 posts
David Kogan
@davidkogan_
21. interp, alignment, and ML efficiency researcher. tryna make ai smarter, cheaper and safer. Building @saferintel.
the matrices Katılım Şubat 2017
452 Takip Edilen62 Takipçiler
1/7 Releasing my first blog post, written over the past few days in spare moments at ICML.
I’ve long been interested in test-time training and looped models, and finally had time to play with their intersection:
“Loop deeper, or adapt?”
alvinzh04.github.io/blog/looped-tt…
English
@ZAlvin39105 Great work Alvin, some of my work involves looped models so I'll be taking a good look later! Abstract already caught my eye. test-time adaptation in recurrent transformers... very, very interesting.
English
@alogictom @SahinLale It isn't confirmation but it is absolutely an indication, and a big step in that direction. I'm working on this too and I wholly believe in the "local agi via low bit" future.
English
is this the confirmation that whatever current dgx spark/ strix/ mac studio that exists in the near future will basically be able to run agi locally?
and hardware just becomes infinitely inflating?
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.
English
@thsottiaux You deserve those 8M. Incredible model, incredible launch transparency, incredible customer treatment.
Thanks Tibo!
English
Hello. We have reached 8M active users across Codex and ChatGPT Work.
We are once again resetting the usage limits for all. And we continue to not have the 5h rate limit as well, allowing everyone to explore the boundaries of GPT-5.6 Sol and discover how ambitious you can be.
See you tomorrow for more updates on our growth!
Sam Altman@sama
5.6 sol growth is insane. the inference team has done heroic work to be able to support demand. we are going to move mountains to continue to scale, but it is possible there are some hiccups soon.
English
@HessianFree @AkhtiamovDanil Which would you say is the more promising format long term, the one PrismML is putting its heart and soul into, ternary or binary?
English
We did a thing.
We just launched Bonsai 27b Binary and Ternary models based on Qwen 3.6. Apache 2.0
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.
English
@sama As long as the codex team keeps being as open about everything as they have been, we will stick along for the ride! You've all been incredible keeping the public informed about every hiccup since and surrounding launch, and GPT-5.6 has absolutely earned its fanfare.
English
5.6 sol growth is insane.
the inference team has done heroic work to be able to support demand.
we are going to move mountains to continue to scale, but it is possible there are some hiccups soon.
English
@pashakho @HessianFree Immediate follow. You're criminally underfollowed, incredible work.
English
We grew a very big model in a very tiny pot. 🌱
Meet Bonsai 27B; a 27B-class multimodal model that fits on a phone.
🪴 3.9 GB at 1-bit
💻 5.9 GB in ternary
🧠 Reasoning, vision, tool use
🔓 Apache 2.0
Small footprint. Full-grown intelligence.
prismml.com/news/bonsai-27b
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.
English
@SahinLale, Cofounder @PrismML, just hinted at GLM-5.2 1-bit coming soon.
1.21b/wt comes out to ~105GB
That fits on ONE SPADK.
If it too can retain ~90% capability, we are looking at an entirely new Pareto frontier for the capability x efficiency tradeoff.
Yeah, this is big.
Sahin Lale@SahinLale
English
27B parameters. Qwen-3.6 running on a PHONE at 90% capability.
If this scales, and if even more quality can be retained, it fundamentally changes what is possible with memory-bound hardware, and in a ram crisis, that's huge.
The future of AI is efficiency, not pure scaling.
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.
English
@PrismML @ClementDelangue Phenomenal work. The future of AI is in more capable, more efficient, more intelligence-dense models. Scaling can only go so far when the substrate is inefficient. Love to see this and cannot wait to see what you put out next!
English
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.
English
@ishgineering it's absolutely incredible. Give it good instructions (so it doesn't go overboard) and it will blow your mind
Have had it working near 24/7 since it came out
English
@davidkogan_ @JeffSte17327059 It isn't open source, but you will see me talk bullshit about it continually until it is :3
English
@1casie @JeffSte17327059 I can't tell if I love the look or hate it... but I lean towards love it. Esp cuz it renders pages so (mostly?) nicely. Is it open source/public yet?
English
@beffjezos try having it write a skill for how it makes design decisions for animations and UI, then import that skill to codex or your faster coding agent of choice
it's not perfect but it does make a difference
English
Fable still untouchable for taste / animations unfortunately.
Need a faster inference option. It's too slow.
English
prediction: Anthropic will reset usage limits in coincidence with whatever reset(s) OAI does to celebrate 8M
English
@thsottiaux gotta clutch up with that 8M celebration reset 🙏 had sol grinding all night on 3 different projects (it was so worth it this is incredible)
English