Z.ai's GLM-5.2, an openly downloadable model with frontier-level coding and bug-finding skills, has revived the debate over whether open-weight AI needs new rules.
OpenAI's Privacy Filter, a roughly 1.5 billion parameter model under an Apache 2.0 license, detects and masks personal information in text before it leaves a user's device.
Mistral AI has launched Voxtral TTS, a lightweight open-weight text-to-speech model built on Ministral 3B that supports nine languages and can clone a voice from a few seconds of audio.
A Microsoft study finds that a common reinforcement learning method can be turned against a model's own guardrails, undoing safety alignment across 15 language models.
Kimi K2.5 adds native image and video understanding, parallel sub-agent "swarms" and switchable reasoning to Moonshot AI's open-weight model line.
Qwen3-TTS supports ten languages, adjusts tone and emotion from natural-language instructions, and streams audio with latency the team says is under 100 milliseconds.
