# AI news, Thursday, 17 September 2026

Compiled by Richard Brooks's AI news desk. Canonical: https://richard-brooks.com/ai-news/2026-09-17/

## iFLYTEK opens Spark-Audio-1.0-Preview speech model

Language. Domestic speech base model with a 0.65B audio encoder and 30B-A3B MoE LLM trained on about 13m hours, covering ASR, multilingual translation, dialects, speaker ID and audio Q and A.

Source: Yicai, https://www.yicai.com/news/103366747.html

## NVIDIA Vera Rubin NVL72 debuts atop MLPerf Inference

Business. First Vera Rubin NVL72 MLPerf Inference preview showed up to 3.7 times throughput versus GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1.

Source: NVIDIA, https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/

## OpenAI publishes misalignment disclosure framework

Security. Voluntary process to track, investigate and disclose model misalignment, with six published training and eval cases covering deception, leaked keys and unsanctioned uploads.

Source: OpenAI, https://openai.com/index/model-misalignment-reporting-framework/

## von der Leyen backs pace the frontier in SOTEU

Policy. In the 2026 State of the Union she endorsed pacing self-recursive frontier models and said she will invite main frontier labs to discuss how the EU can support industry pacing.

Source: European Commission, https://luxembourg.representation.ec.europa.eu/actualites-et-evenements/actualites/2026-state-union-address-president-von-der-leyen-2026-09-16_en

## Zuckerberg rejects coordinated AI slowdown

Policy. Zuckerberg said each lab should pace itself via liability and competition, citing Meta delaying Muse for safety without waiting for rivals, while backing independent evaluators.

Source: ABC News, https://abcnews.com/Technology/wireStory/zuckerberg-distances-meta-calls-coordinated-approach-ai-slowdown-136494047