The Day in AI

• Ant Group released the weights of Ling-3.1-flash, a 560 billion parameter model with a 1 million token context window.
• Elorian released Glimpse, its first model, which crops images and pulls video frames while it reasons, at $0.60 per million input tokens.
• Meta released Muse Spark Open, an 816 billion parameter open-weight model, and says it leads GLM 5.3 narrowly on JobBench.
• Xiaomi open-sourced MiDashengLM-Spatial, an audio-language model of about 8 billion parameters that locates where sounds come from.
• China's Ministry of Commerce is reportedly considering adding 23 AI technologies to its export control list, according to Bloomberg.

Ant Group built Ling-3.1-flash as a mixture-of-experts model, which means each token is routed through a small subset of the weights. It targets coding, agent workflows and long professional documents that use external tools, and it supports a context window of up to 1 million tokens. Deployment support is in SGLang and vLLM. Ant Group reports that a mixed-precision setup cuts weight memory from about 534GB to 292GB, so a Hopper GPU machine needs four GPUs instead of eight.

Elorian released Glimpse as a research preview with API access, priced at $0.60 per million input tokens and $3 per million output tokens. Most multimodal models look at an image once, turn it into tokens and then reason in text. Glimpse can go back to the source in the middle of its reasoning to crop a region, pull specific video frames or run code to measure and compare, and Elorian calls this a visual scratch space. The company says a small team trained it on a few hundred GPUs and that it is much smaller than the frontier systems it compares itself with.

Meta released Muse Spark Open on 15 October, its largest open-weight model, with about 816 billion total parameters and 42 billion active per token under a restricted license. It takes text, images, audio and video and outputs text. Meta reports 61.8 on JobBench against 61.4 for GLM 5.3, and 76.4 on Terminal-Bench 2.1, where Kimi K3 scores 85.0. The model card lists an Artificial Analysis Intelligence Index of 39, below GLM 5.3 at 45. Meta says it worked with Apple so the model runs locally with MLX on two M5 Ultra Mac Studios.

Ling-3.1-flash and Muse Spark Open came out within a day of each other and both have 1 million token context windows. Neither has independent benchmark results yet.

• Xiaomi open-sourced MiDashengLM-Spatial under Apache 2.0, which reasons about the direction, distance and movement of sounds from binaural audio.
• IBM released Granite-Switch-4.1 in 8B and 30B sizes, combining chat, a safety adapter and speech recognition in one model and one endpoint.
• OpenAI added native Decisions integration to GPT-Live on 15 October, so apps can act while the user is still speaking.
• OpenAI has reportedly opened prompt-to-voice to all API customers, according to its changelog, which builds a reusable GPT-Live voice from a text description.
• OpenAI introduced the Rosalind Academic Plan, which gives eligible academic labs up to 30 free Business Standard seats for 12 months.
• Apple says Apple Intelligence will enter beta in China in late October using Alibaba's Qwen models, and a December launch is reported.
• Nathan Chen Guangyu, a 17-year-old researcher at Moonshot AI, said on 15 October he has left the company and is moving to the United States.
• China's Ministry of Commerce is reportedly weighing export controls on 23 AI technologies, including large language models and chip designs, according to Bloomberg.
• Taiwan plans to curb AI server and chip exports to China and make shipping restricted hardware a criminal offense, Bloomberg reports.
• CERT-In, India's cyber agency, said AI agents broke into Indian organisations' systems and read confidential files undetected, Reuters reports.