Open Source AI

Hugging Face has moved Safetensors into the PyTorch Foundation, turning the default open-model weight format into vendor-neutral infrastructure for the wider stack.

NVIDIA's Nemotron Coalition is an attempt to turn DGX Cloud, post-training, and partner coordination into the shared foundation for frontier open AI.

Z.ai's MIT-licensed GLM-5.1 and GLM-5.1-FP8 checkpoints hit Hugging Face with vLLM and SGLang support, so the benchmark story now has to survive deployment reality.

Netflix has released VOID, a video-editing model that removes objects and their physical aftermath, but using it still means masks, two passes, and 40GB-plus VRAM.

Gemma 4 is not one model. Google's E2B, E4B, 26B A4B, and 31B force real choices about memory, latency, audio, context, and reasoning quality.

Gemma 4's real launch is the stack around it: Apache 2.0 weights, AICore, AI Edge Gallery, LiteRT-LM, and day-one local-agent support.

Microsoft Agent Framework is now Microsoft’s shared successor for AutoGen and Semantic Kernel, with migration guides and RC signals making the shift explicit.

H Company’s Holo3 pairs a benchmark-leading gated flagship with an Apache 2.0 smaller model, giving developers a real open-weight foothold in computer-use AI.

IBM's Granite 4.0 3B Vision is a compact Apache 2.0 model for charts, tables, and form fields, giving document AI an open enterprise wedge.

GitHub Squad turns multi-agent coding into a concrete repo-native pattern, with shared decision files, separate reviewer agents, and team memory that lives in git.

Lark CLI, DingTalk Workspace CLI, and WeCom CLI show workplace suites turning messages, docs, meetings, and tasks into agent-ready action layers instead of just prettier chat boxes.

Voxtral TTS gives builders more control over voice agents, but Mistral's CC BY-NC license keeps the self-hosted commercial path behind a gate.