Nvidia confirmed the $12.93B Hugging Face acquisition on Sept 3, 2026, with written openness commitments. What changes, what to watch, and how to insulate your projects.
Tencent's Hy4 preview ships 770B parameters under Apache 2.0 with a 1M-token context and verified API pricing near $0.83 per million input tokens — the full guide.
Z.ai released the full GLM-5.3 weights — 753B parameters, 755.7 GB — free on Hugging Face. The benchmarks, license terms, and how to run it, explained.
Hugging Face's Summer 2026 report: 2.96M open models, Chinese labs peaking at 2.78T parameters, AMD and NVIDIA top publishers, and 1.5% of repos taking 99.2% of downloads.
LTX-2.5 is a 22B open-weights model by Lightricks featuring native multishot generation with full continuity of characters, lighting, and audio across cuts.
MiniMax-H3 is a 33B open-weights model that generates high-quality video with natively synchronized stereo audio from text and images. Runs locally via ComfyUI.
Unsloth launches with NVIDIA the first desktop app to train, fine-tune, and run AI models locally on Windows, macOS, and Linux with full fine-tuning support.
Cohere releases Cohere Transcribe Arabic — the most accurate open-source Arabic ASR model, outperforming Whisper v3 Large by 11 points and handling Arabic dialects and Arabic-English code-switching.
Start typing to search
