Hugging Face

The Allen Institute for AI has decided that humanity's greatest bottleneck is the time spent formatting corporate summaries, releasing AstaBrief as an open-source report-generation model designed for rapid document creation. * Open-sourcing models sounds noble until you realise you still have to host, fine-tune, and clean up after them.

AllenAI has released Olmo-core 3, an open-source training infrastructure tailored specifically for large Mixture of Experts (MoE) models. By open-sourcing the complex scaffolding required to wrangle these sprawling architectures, they are inviting the broader community to shoulder the computational misery of frontier model training. * Olmo-core 3 provides open scalable training

Hugging Face has introduced the Open TTS Leaderboard, a scalable platform designed to evaluate multilingual text-to-speech and voice cloning models. Because the world was clearly crying out for yet another benchmark to obsess over, this initiative promises to bring some much-needed order to the chaotic landscape of synthetic voice generation.

Nvidia has released Nvidia Kumo Tabular, a new model architecture aimed at solving the unglamorous world of tabular prediction. While the rest of the tech industry chases multimodal generative AI, Nvidia is quietly trying to optimize enterprise spreadsheets and boring business forecasting. * Boring is profitable: This model targets tabular data,

Hcompany has released Holo4, an open-source framework designed to power generalist computer-use agents that can supposedly interact with software just like a human. As the race to build autonomous desktop assistants continues, this latest offering attempts to bridge the gap between static language models and active software execution. * Open-source framework

Hugging Face has released tokenizers v1, bringing long-overdue stability and performance metrics to the foundational text-processing libraries used across the machine learning ecosystem. * Standardisation arrives years after the industry adopted the underlying architecture, proving that speed to market matters more than version numbers. * New benchmarking data confirms marginal speed gains

In a move that surprises absolutely no one paying attention to open-source consolidation, Jun Kim, the creator and maintainer of oMLX, has officially joined Hugging Face to bolster support for the Apple MLX community. * Apple silicon developers get centralized support for oMLX under the cosy corporate umbrella of Hugging Face.

Hugging Face has quietly integrated llama.cpp quantization directly into the Transformers library, bridging the gap between heavy research models and local execution hardware. Developers no longer need to jump through endless conversion hoops just to run compressed models efficiently on consumer-grade machines. * Local deployment of heavy models just got

Liquid AI has released LFM2.5-VL-DSpark, promising to speed up vision-language models for developers drowning in multimodal data. Because what the world really needs is another benchmark-chasing architecture to evaluate. * Liquid AI claims architectural efficiency improvements over standard vision-language models. * Developers can access and test the new model directly via

Nvidia has rolled out a guide on leveraging Nvidia Warp and MjWarp to accelerate robotics simulation and learning workflows, offering developers a way to bypass the notoriously glacial pace of training virtual machines. * Faster simulation means your virtual robots can fall over millions of times before breakfast. * By shifting heavy