Sabitlenmiş Tweet
Discover your @huggingface profile into a huggimon holo Card at huggimon.co !
- Discover your holo card
- Fill in your binder with rarest cards
Catch'em all ! Become the best huggimon Trainer ! 🤗
English
【T】【A】【M】【S】
1K posts
【T】【A】【M】【S】
@tamsi_besson
AI & Builder, believe in local AI ! Building : - https://t.co/BPTKILpwfI a HUGE QA solution(try the free version it's sick) - https://t.co/n0J41nn0Dp huggingface profile pokemon-style
Paris, France Katılım Mayıs 2017
96 Takip Edilen131 Takipçiler
【T】【A】【M】【S】 retweetledi
What is Sliding Window Attention?
Sliding Window Attention (SWA) is a local attention pattern that restricts each token to attending only within a fixed-size neighborhood instead of the full sequence. This reduces attention and KV-cache memory for long-context models, while periodic global-attention layers can preserve broader context.
Learn more here, alongside the 208 papers that cite it: paperswithcode.co/methods/slidin…
elie@eliebakouch
fun fact, literally all the open models from western "big labs" (with a lot of ex close lab people), so thinking machine, microsoft MAI, gemma, gpt-oss have one thing in common: they ALL use sliding window attention 🙂
English
【T】【A】【M】【S】 retweetledi
【T】【A】【M】【S】 retweetledi
Kimi K3 is about to have America rethinking its life choices tomorrow.
What do you do with a Chinese model that’s approaching SoTA, but 50x cheaper?
Oh, yea. And it’s open-source.
Kimi.ai@Kimi_Moonshot
English
@TheAhmadOsman It’s as expensive/cheap as sonnet.. depends on the benchmark
English
【T】【A】【M】【S】 retweetledi
Bonsai 27B running locally on an iPhone in Atomic Chat!
Bonsai is the first 27B-class model that fits on a phone. @PrismML built it on Qwen3.6 27B with 1-bit weights. It takes 3.9GB instead of 54GB and keeps ~90% of the benchmark scores.
Available now on iPhone and Android
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
【T】【A】【M】【S】 retweetledi
11 hours ago kimi update --> v0.25.0
2 hours ago kimi update --> v0.26.0
And K3 is here!!!!!!
English
【T】【A】【M】【S】 retweetledi
Introducing a new KV format: VBR (Variable Bit Rate)
VBR dynamically quantizes your KV cache layer-by-layer as your session grows, giving you the highest quality possible within your VRAM constraints. This is my dream format.
Pushed to master. Available now. 1/15 🧵
English
English
@Tech2Wild Can't wait to see what the 1-bit quant of this small model can do
English
【T】【A】【M】【S】 retweetledi
Pay Attention !
If you dig into the HF page, there's also Inkling-Small hiding out: 276B params, 12B active, and it actually outperforms the big model on some benchmarks.
Not a formal launch, just quietly sitting there. Easter egg find 🥚
Open weights, Apache 2.0, fine-tune on Tinker.
Thinking Machines@thinkymachines
Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
English
【T】【A】【M】【S】 retweetledi
We collabed with Thinking Machines to bring dynamic Unsloth 1-bit GGUF quants for Inkling to everyone! 86% smaller (270GB vs 1.9TB) yet it retains 74.2% of top-1% accuracy!
We added vision and audio support as well! GGUFs and more at huggingface.co/unsloth/inklin…
Unsloth AI@UnslothAI
You can now run Thinking Machines Inkling! Inkling is a 975B open model with image, audio and 1M context support. We quantized Inkling to Dynamic 1-bit (-86% size) and retained 74.2% of top-1% accuracy. Run on 280GB. Guide: unsloth.ai/docs/models/in… GGUF: huggingface.co/unsloth/inklin…
English
@LocallyAIApp @PrismML I havé an iPhone 17.. so i can’t see the 1-bit bonsaï 27b on LM studio
English
Meet 1-bit Bonsai 27B, the latest model from @PrismML and the first 27B-class model to run on a phone.
This model, available on the iPhone 17 Pro, iPhone Air, and select iPads, pushes the limits of intelligence density. Powered by MLX.
Available now.
English
【T】【A】【M】【S】 retweetledi
@LysandreJik It seems awesome but i think i don't get the real revolution here, anyone to explain ?
English
Thinking Machines' Inkling is out: first ever open and large (1T), text, image and audio in, text out.
One thing I find quite striking is how much easier accelerating models has become.
We replaced the model's causal Conv1D with the `causal-conv1d` kernel. One line changed, +4% tokens per second.
We then replaced its attention implementation with FlashAttention-4. Another single change, another +11%.
That's a total throughput improvement of about 15%, without changing the model architecture or retraining anything.
And we haven't touched the MoE layers yet, which should offer some of the largest opportunities for optimization.
This is what makes the `kernels` effort so exciting to me: enabling model developers to find and integrate the best optimized kernels almost as easily as they would change a model configuration.
Kernel developers can focus on building the fastest implementations. Model developers can test and adopt them quickly. Everyone using the model benefits.
Two light updates for 15% on a 1T model is a pretty good start.
English
Look what i've built :
a full course chapter by chapter with link to videos and courses on github.com/Tamsi/ai-lab/b…
That's what i'm using to learn maths behind AI, ML, DL, RL, CNN, RNN, etc...
Some courses are from @karpathy and @DeepLearningAI
【T】【A】【M】【S】@tamsi_besson
I'm searching for a good online AI Diploma, i saw : - Woolf x Udacity -> udacity.com/masters-artifi… - CU Boulder x Coursera -> coursera.org/degrees/ms-art… If anyone has other proposals, i'm interested 👀
English
【T】【A】【M】【S】 retweetledi
“L’IA c’est de la merde, ça hallucine, ça remplacera jamais un humain.”
Signé: un mec qui l’a testée 10 minutes en 2024 et qui a formé son opinion pour toujours.
Un jugement daté d’une version morte, appliqué à une techno qui double tous les 3-6 mois.
Critiquer Claude 2026 avec ton souvenir de GPT 2024, c’est critiquer l’iPhone en se basant sur le Nokia 3310.
Le pire c’est que ces mecs sont souvent fiers de leur constance. “J’ai toujours dit que…” Oui. C’est exactement le problème. T’as figé ton avis pendant que l’objet de ton avis mutait.
Une opinion sur l’IA a une date de péremption de 90 jours. Et encore…
Français
@jun_song 3T model is huge, can't use it locally on mac or pc even in 2-bit
English
