The day in AI

Aleph Alpha, the Heidelberg lab that sells to European governments and industry, put Kolibri's weights on Hugging Face under Apache 2.0. A mixture of experts model sends each token through a small subset of its expert sub-networks. So only about 3B of Kolibri's 78B parameters do work on any one token, and it runs much cheaper than a dense model of the same size. Aleph Alpha says the context window goes up to 1M tokens. A third-party write-up puts the active count closer to 3.5B.

Kolibri was trained from scratch. Aleph Alpha first tested its training pipeline on Kolibri Origin, a smaller model with 30B total parameters, 3B active and a 65k token context, and then scaled up. The company tuned Kolibri for German, reasoning, math and agentic tasks, aimed at regulated buyers in public administration, industrials and aerospace.

On its own German evaluation, Aleph Alpha reports a score of 70.8 for Kolibri against 69.8 for Qwen3.5 35B-A3B, a mixture of experts model with a similar active size. The dense Qwen3.8 27B scores 79.9 on the same evaluation, so Kolibri is ahead of its size class and well behind the stronger dense model. No independent German benchmark results had been published by that day.

Meta released open-source firmware and a Linux SDK on 2 October so developers can build their own hardware around Muse, its assistant. Until then Muse ran only in Meta's apps and on the upcoming Muse Charm device. The code is Apache 2.0, but the tokens that connect a device to Muse's cloud are free only for personal, non-commercial use. A developer can give away up to 50 devices and can't put a token in anything sold or publicly advertised. Meta supplies no hardware, and Codersera lists suitable boards from a $12 Waveshare ESP32-C6 to a $99 Seeed reTerminal. US subscribers also get a free Muse Home Link USB-C dongle.

What isn't yet known is whether Meta will open commercial token licences, which would decide if the SDK reaches products or stays with hobbyists.

• David Robinson, who led drafting of OpenAI's safety reports, published an essay in The Atlantic explaining his 2 October resignation, according to reports, saying he oversaw reports on 12 frontier launches and plans to work on safety from outside.
• FTC has opened an investigation into OpenAI and Anthropic over possible risks to consumers, according to several outlets.
• California's attorney general subpoenaed OpenAI over cyber incidents involving AI agents, as reported by Tom's Hardware.
• US Government plans to propose an emergency AI notification system with China, according to Axios.
• Google has reportedly held back public release of a new model over cyberattack concerns, according to The420.in.
• arXiv has started limiting researchers to two submissions a month, as reported by Cybernews and others.
• Jay Clayton, Director of National Intelligence, was named to lead a new White House AI task force, according to CNBC.
• Anthropic will reportedly spend $100M on an academy to train 10,000 AI engineers.
• US Department of Justice charged a California man in an alleged $300 million Nvidia AI server smuggling case, according to reports.
• xAI won a ruling from a US appeals court that blocks Minnesota's ban on AI nude images of real people, according to India Today.