Francesco Visin

180 posts

Francesco Visin

Francesco Visin

@FrancescoVisin

Senior research Scientist @DeepMind. Co-organizer @M2LSchool. In a previous life @Mila_Quebec. All opinions are my own, some of the ideas as well. He/him.

Katılım Ekim 2011
263 Takip Edilen472 Takipçiler
Sabitlenmiş Tweet
Francesco Visin
Francesco Visin@FrancescoVisin·
I couldn't be more proud of this tiny giant that is a great addition to the Gemma family!! I can't wait to see what the community will build with this powerful model!
Omar Sanseviero@osanseviero

Introducing Gemma 3 270M 🔥 🤏A tiny model! Just 270 million parameters 🧠 Very strong instruction following 🤖 Fine-tune in just a few minutes, with a large vocabulary to serve as a high-quality foundation developers.googleblog.com/en/introducing…

English
0
0
14
928
Francesco Visin retweetledi
Omar Sanseviero
Omar Sanseviero@osanseviero·
Introducing the MedGemma Impact Challenge Use MedGemma and other open models to build human-centered applications in the medical space. Over $100k in prizes. Time to build. kaggle.com/competitions/m…
English
14
66
613
34.8K
Francesco Visin
Francesco Visin@FrancescoVisin·
We brought function calling to your toaster! 🤩 🤏 Just 270m: fits on your phone, tablet - and probably your doorbell 💫 Lightning Fast, low latency 🫵 Bring Your Data: built with finetuning in mind, make it your own 🔑 Safe & Private 🤗 Huge Ecosystem Support
Omar Sanseviero@osanseviero

Introducing FunctionGemma 🤏270m model for function calling 📱can run in your phone, browser or other devices 🤖designed to be specialized for your own tasks huggingface.co/collections/go…

English
0
0
3
75
Ivan Morgillo
Ivan Morgillo@hamen·
I tried to replace Qwen2.5 7B with the new FunctionGemma. I failed. The model is very small and it panics on every basic prompt. I know that 270M vs 7B is unfair, but I hoped it would fight a bit harder. Back to Qwen2.5 for now.
English
2
0
1
479
Francesco Visin retweetledi
M2L school
M2L school@M2lSchool·
All recordings are now live! #M2L2025 talk-sharing journey is officially here. 👉 Watch the full playlist: youtube.com/playlist?list=… 🚀 Stay tuned — more updates coming soon!
M2L school tweet media
English
0
2
10
5.1K
Francesco Visin
Francesco Visin@FrancescoVisin·
Heading to yet another edition of @M2lSchool, this time in Split, Croatia! 🇭🇷 Super excited to spend some days surrounded by passionate students and outstanding lecturers! 🤩 If you are at M2L this year don't be shy, come say hi! 👋 #AI #ML #education
English
0
0
1
143
Francesco Visin retweetledi
Sundar Pichai
Sundar Pichai@sundarpichai·
Introducing EmbeddingGemma, our newest open model that can run completely on-device. It's the top model under 500M parameters on the MTEB benchmark and comparable to models nearly 2x its size – enabling state-of-the-art embeddings for search, retrieval + more.
English
199
520
7.5K
534.8K
Francesco Visin
Francesco Visin@FrancescoVisin·
Gemma 270M can be finetuned and deployed in production *in minutes*!
Mr. Ånand@Astrodevil_

I just fine-tuned @GoogleDeepMind's "gemma-3-270m" and it's ready for secure deployment in production! → Imported and fine-tuned using @huggingface transformers on M4 → Packaged and versioned with @Kit_Ops → Pushed to @Jozu_AI Hub for production use and deployments All within just a few minutes🔥 "gemma-3-270m" is lightweight, powerful, and designed for production-grade systems

English
0
4
11
757
Francesco Visin retweetledi
Paul Couvert
Paul Couvert@itsPaulAi·
Google has released a new open source model... That runs on just 0.5 GB of RAM. Yes. You can fine-tune it for free to make it better than the giant models at your tasks. Quick steps to fine-tune Gemma 3 270M below
Paul Couvert tweet media
English
74
406
3.5K
441.4K
Francesco Visin retweetledi
Adrien Grondin
Adrien Grondin@adrgrondin·
Gemma 3 270M (8-bit) at ~200 tk/s on iPad Air M3 with MLX
English
20
23
299
30.2K
Francesco Visin
Francesco Visin@FrancescoVisin·
@kchonyc Hey Cho - long time! We designed it as a generalist model with the goal to be highly effective across a wide range of tasks after fine-tuning. Used out of the box, small tweaks to the prompt usually go a long way in fixing minor quirks. LMK if that helps!
English
0
0
2
263
Kyunghyun Cho
Kyunghyun Cho@kchonyc·
hmm .. gemma3:270m spuriously repeats too often.
English
5
0
9
4.9K
Federico Baldassarre
Federico Baldassarre@BaldassarreFe·
Say hello to DINOv3 🦖🦖🦖 A major release that raises the bar of self-supervised vision foundation models. With stunning high-resolution dense features, it’s a game-changer for vision tasks! We scaled model size and training data, but here's what makes it special 👇
Federico Baldassarre tweet mediaFederico Baldassarre tweet mediaFederico Baldassarre tweet mediaFederico Baldassarre tweet media
English
40
261
1.9K
223.9K
Francesco Visin
Francesco Visin@FrancescoVisin·
@eliebakouch @mciccone_AI Also, since the embedding matrix is effectively a lookup table, it can be implemented efficiently on CPU if needed. But I agree with you, there is some room to squeeze even more performance out of the tiny model size with smaller vocabularies.
English
1
0
1
351
Francesco Visin
Francesco Visin@FrancescoVisin·
@eliebakouch @mciccone_AI We considered it, but a consistent tokenizer across the Gemma family makes it easier to develop ideas rapidly at small scale and test them on more powerful models with minimal effort.
English
1
0
1
93
elie
elie@eliebakouch·
Super impressive model, especially since there is only 100M non embedding param. 😮 Would be super curious to know if they change anything in the training to get better perf at this scale compare to the 1B+ models?
elie tweet media
English
6
7
77
20.3K
Omar Sanseviero
Omar Sanseviero@osanseviero·
Introducing Gemma 3 270M 🔥 🤏A tiny model! Just 270 million parameters 🧠 Very strong instruction following 🤖 Fine-tune in just a few minutes, with a large vocabulary to serve as a high-quality foundation developers.googleblog.com/en/introducing…
Omar Sanseviero tweet media
English
118
317
2.6K
727.5K
Francesco Visin
Francesco Visin@FrancescoVisin·
@respectorr69 @osanseviero Extremely good in our tests! But keep in mind that this model is mainly meant to be finetuned, rather than be used out of the box
English
1
0
0
96
Respect
Respect@rrespectorr·
@osanseviero How good is it for structured output? Like turning a sentence into JSON while also extracting relevant information like actions and locations?
English
2
0
0
390
Francesco Visin
Francesco Visin@FrancescoVisin·
@eliebakouch @mciccone_AI Thank you!! We've built on the fantastic Gemma3 recipe, with tailored tweaks on e.g., data mixtures, optimized for the use cases of very small models.
English
1
0
3
151
Francesco Visin
Francesco Visin@FrancescoVisin·
@M2lSchool 2025 applications are open!!✨ Diversity is a core value for us. We actively encourage students from diverse backgrounds to apply! Please share widely 🙏 #AI #ML #SummerSchool #DEI #m2lsummerschool #m2l #WomenInSTEM #WiML
M2L school@M2lSchool

We are excited to announce we have opened the applications for the 5th edition of the Mediterranean Machine Learning (M2L) summer school in September 2025! This year, the school will take place at University of Split (@HrUnist) in Split, Croatia. Apply at m2lschool.org/application

English
1
0
4
238

Keşfet

@hamen @M2lSchool @kchonyc @BaldassarreFe @eliebakouch @mciccone_AI @JLopez_160 @osanseviero