When using the terminal, I often find myself typing the same few commands over and over.

What if instead, you could yell them at your computer?

Introducing:🗣🖥️ Talk To Your Terminal, an extremely personal human-computer interface.

github.com/aukejw/ttyt

We are back at #ICCV2025 for the second workshop on Hyperbolic and Hyperspherical Learning for Computer Vision.

We will have 2 tracks: one for new research (to be published in proceedings) and one for recently published works.

Deadline: June 1st
Link: sites.google.com/view/beyonde...

Qwen3 4bit and 8bit models are now available, and they are fast!

For the 8B model, I'm seeing 45-55 tokens/sec with mlx_lm, and 25-40 tokens/sec with LM Studio depending on context length.

(do take LM Studio peak memory with a grain of salt, I'm sure it's not constant)

I've been benchmarking Gemma 3 models on my M4 Pro to see how fast MLX is. The answer is: very!

Below: generation with the 12B model (4 bit quantized).

However, I found it hard to find out which models my Macbook could handle, and how fast I *should* expect inference to be.

One week left to submit your work to eLVM2!

We're looking for reviewers for the 2nd edition of our CVPR workshop on efficient large vision models (eLVM).

If you're a researcher working on efficiency of foundation models, and are interested in reviewing, let me know by March 14th!

forms.gle/Ra1dDwmzjxyT...

🚀Call for papers🚀

We're organizing the 2nd workshop on Efficient Large Vision Models at CVPR 2025.

Welcoming all submissions on efficient (visual) foundation models!

⏰ Submission deadline: March 21
🌐 sites.google.com/view/elvm/

Turn a video of my face (or Jens, Guillaume,..) into anything at our NeurIPS booth -- running a video diffusion model on a mobile phone ->