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.

We're opening a Hugging Face office in Tokyo! Our goal: help open-source AI develop in Japan and grow the local community. Let's meet! ハギングフェイスの東京オフィスがオープンしました！ 私たちの目標は、日本におけるオープンソースAIの発展を支援し、ローカルコミュニティを育てることです。ぜひお会いしましょう！

Transformers v5.4.0 is live with 13 new models! Key highlights: • VidEoMT: 160 FPS video segmentation • Document AI suite: UVDoc, SLANeXt, PP-OCRv5 • PI0: Vision-language-action model for robotics • Mistral 4: Unified instruct/reasoning model • Jina Embeddings v3: Multi-task text embeddings • CHMv2: Forest canopy height mapping ~10 new model architectures contributed by the PaddlePaddle team 🤯 Plus major performance improvements and Flash Attention 4 support!

PyTorch Conference Europe is happening next month, and many of us from the Hugging Face team will be there We'll be talking about Kernels, Transformers, Audio models, processors, Diffusers, Transformers.js among others We'll be sharing a few sneak peaks of what's coming next 🤫