Microsoft and Meta reveal giant models as FTC and US states target OpenAI

Microsoft finishes pre-training MAI Base 2

On 15 October, Microsoft AI said it had finished pre-training MAI Base 2 and that the model is now in post-training. It's a mixture-of-experts model with 2.8 trillion parameters in total, of which 80 billion are active. Training ran for 38 days across August and September on 16,000 GB200 GPUs, using a 50 trillion token mixture. Vision is trained in from the start, using 4.5 trillion unique vision-language tokens.

Microsoft says MAI Base 2 beat every open-weight model it tested on its own held-out evaluations. It also says the model used 4.2 to 12.3 times less pre-training compute than DeepSeek-V4-Pro-Base. Microsoft says all of the training data was processed in-house with no distillation from other models, and it intends to license the weights to trusted partners through Foundry.

What it means for NZ organisations: Azure customers may soon get a frontier model that Microsoft built itself. If you're close to signing a long platform deal, it's worth seeing the Foundry licensing terms first.

Meta publishes Muse Spark Open weights

Meta released Muse Spark Open on 15 October. The model is multimodal from the ground up and aimed at coding, software engineering and long agent tasks. It has 816 billion parameters, 42 billion of them active, and a context window of one million tokens. Meta says two Apple M5 Ultra Mac Studios running MLX can hold the whole model. SGLang, vLLM, llama.cpp and Hugging Face Transformers support it too.

The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Groq, Fireworks AI, Together AI, OpenRouter, Baseten, Modal and Nebius offer hosted versions. The Muse Model License has limited commercial terms, and many developers won't pay anything to use it.

What it means for NZ organisations: Teams that have to keep sensitive code or data on their own hardware now have a strong local option. Have the licence checked before you build anything you'll sell.

Nvidia releases Nemotron 3.5 Super VL

Nvidia launched Nemotron 3.5 Super VL on 15 October. It's an open 120 billion parameter vision-language model with 12 billion active parameters, built for agents that work with PDFs, web pages, charts, tables and scanned text. Citing the Artificial Analysis index, Nvidia says it's five times faster than the next best open agentic model in its class. It's available on Hugging Face, OpenRouter, ModelScope and Nvidia NIM.

What it means for NZ organisations: If your workflows pull data from invoices, forms or reports, compare this model with your current paid API on cost and speed.

Eight US states sue OpenAI over ChatGPT safety

On Wednesday 14 October (US time), attorneys general in Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed separate lawsuits against OpenAI. They say the company misled users about how safe ChatGPT is, in breach of state consumer protection law. They also argue that a chatbot designed to be agreeable can reinforce delusions and encourage dangerous behaviour. The states want court orders forcing major changes to the product. Florida filed the first state case in June.

OpenAI spokesperson Drew Pusateri said the company has built "industry-leading age-specific protections" into ChatGPT. OpenAI also pointed to parental controls and break reminders, and said no system is perfect.

What it means for NZ organisations: If you run a customer-facing chatbot, write down the safeguards you use, especially where young people might use it. Safety claims you can't back up could cause trouble under the Fair Trading Act.

FTC sends investigative demands to OpenAI and Anthropic

The US Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic on the same Wednesday. Axios first reported them on 15 October, and an FTC spokesperson later confirmed them to CBS News. Each demand runs to more than 50 pages and asks for records on security incidents, safeguards, marketing and privacy policies. The FTC also plans to make senior executives testify, and it'll send demands to third-party AI organisations later. OpenAI said it would "engage constructively with the FTC". The agency had announced its investigation about two weeks before.

What it means for NZ organisations: Ask your AI vendors how they record and disclose security incidents, and use their answers in your privacy and risk assessments.

White House AI task force meets for the first time

The White House AI task force held its first meeting on Thursday 15 October and adopted a 120-day charter. It asked the President's Council of Advisors on Science and Technology to assess risk in three areas: cybersecurity and critical infrastructure, biosecurity and dual-use applications, and model control and accountability. Vice chair Emil Michael, the US Defense Undersecretary, is writing protocols that would require labs to report breaches and hacks to the government. It isn't yet clear how they'd be enforced.

What it means for NZ organisations: US reporting rules for labs could end up in the incident clauses of your vendor contracts. Check those clauses when contracts come up for renewal.

Bank of America adds staff to fix flaws found by Mythos

Bank of America chief executive Brian Moynihan spoke on the bank's third-quarter earnings call on 14 October. He said the bank brought in 1,000 contractors during the second quarter to work alongside about 3,000 technical staff on vulnerabilities found by Anthropic's Mythos model. He put the cost of that work and anti-money-laundering remediation together at "a couple of hundred million plus", without giving separate figures.

On 16 October, 420 scientists, including Geoffrey Hinton, Yoshua Bengio and Stuart Russell, published an open letter. It asks developers to keep cyber-capable models off the internet at first, and calls for an international body that would track containment failures.

What it means for NZ organisations: AI scanners can find flaws faster than teams can fix them. Set aside patching staff and budget before you start AI-driven security testing.