🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Bonjour!

PaliGemma and Gemini fueled the new vision modality of Gemma 3. It is so great to work at the crossroads of so many amazing teams!

The MLX community never ceases to amaze me.

Kudos to our colleagues and external partners that work relentlessly in the shadow to ship Gemma across platforms! Gus (🤖🧠+🐍+🥑🗣️) Omar Sanseviero Philipp Schmid and many others

Gemma 3 would not be as good without Aishwarya Kamath amazing work!

That's why I love shipping open models so much

Gemma everywhere!

Insane to see all the complex pieces that have come together to produce this beast. The teams behind it are just incredible.

Cool long context demo presented by Tatiana Matejovicova from Gemma team: youtube.com/watch?v=TAlkGT…

Follow the guide by Ravin Kumar from Gemma team on how to pick model sizes: youtube.com/watch?v=qcjrdu…

We spend a lot of extra time to cook almost lossless quantized Gemma models. Kudos to Edouard Yvinec and Morgane Rivière for their work!

Putting open models out there is hard and releases often come with implementation bugs. For Gemma 3, a dedicated release team worked for weeks to avoid these issues. So let s wait before judging llama 4 performances....

🐬Gemma blog.google/technology/ai/…

QAT models also available on llama.cpp and MLX!

Introducing Qwen3! We release and open-weight Qwen3, our latest large language models, including 2 MoE models and 6 dense models, ranging from 0.6B to 235B. Our flagship model, Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, general

Somehow my twitter thread is not saturated with posts about Qwen 3 🤨 anyway, congratulations Qwen on this release!

we love lmsys and its impact to our field, and thats why we need critical work, like cohere paper to keep improving it. Equally, it is important to clarify that some numbers in Cohere work are inaccurate, e.g., we sent only 1 gemma3 model to lmsys before release ("zizou-10").

It is critical for scientific integrity that we trust our measure of progress. The lmarena.ai has become the go-to evaluation for AI progress. Our release today demonstrates the difficulty in maintaining fair evaluations on lmarena.ai, despite best intentions.

What a week for the open community: Qwen3, DeepSeek and Helium 1!

Lot of fun working on this open release. It is packed with innovations to get the best on-device model out there.