# AI news, Friday, 4 September 2026

Compiled by Richard Brooks's AI news desk. Canonical: https://richard-brooks.com/ai-news/2026-09-04/

## Anthropic, OpenAI and xAI hit elevated errors together

Tools. Three frontier providers degraded in one afternoon. Treat which model you call as a runtime fallback, not a build-time choice.

Source: Anthropic, https://status.claude.com/incidents/461yvfrzpwtt

## Google ships Gemini 3.8 Flash and Flash Cyber

Models. Cheap Flash-tier model aimed at long-horizon coding, with a restricted Cyber variant and Fairwind Program via CodeMender.

Source: Google, https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/

## MBZUAI releases Apache-2.0 K2 Horizon model family

Models. Six open models from 0.9B to a 375B-A23B MoE with 512K context and vLLM/SGLang recipes for agent backends. The 375B card is English-only.

Source: MBZUAI, https://huggingface.co/IFM/K2-Horizon-375B-A23B

## Meta releases Muse Spark 1.3 frontier agent model

Communication. Trained to ask clarifying questions, call the user when stuck, and confirm before consequential actions, with fewer tool calls and stronger prompt-injection resistance than 1.2.

Source: Meta, https://research.meta.ai/blog/introducing-muse-spark-1-3

## Microsoft open-sources VibeVoice-ASR-Streaming 7B

Language. Open-weight 1.5B and 7B models produce speaker-attributed transcripts as audio arrives, with custom hotwords and 10 languages.

Source: Microsoft, https://huggingface.co/microsoft/VibeVoice-ASR-Streaming-7B

## NVIDIA agrees to acquire Hugging Face for $12.93bn

Business. A chip vendor will sit atop the default open-weights and eval distribution layer. Close expected H1 2027, subject to regulators.

Source: NVIDIA, https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/