Microsoft details MAI Base 2 and DeepSeek updates its open Flash model

Microsoft trains a 2.8T-parameter base model in 38 days

Microsoft AI published details of MAI Base 2 on 15 October. It's a mixture-of-experts model with 2.8 trillion parameters in total, and 80 billion of them are active for any given token. The pretraining run used 16,000 GB200 GPUs for 38 days across August and September. The training mixture held 50 trillion tokens, its knowledge cutoff is March 2026, and Microsoft says no outputs from other models were distilled into it.

Microsoft's own figures put it at 4.2 to 12.3 times more compute-efficient in pretraining than DeepSeek-V4-Pro-Base. Microsoft also says it outscored every open-weight model it tried on its held-out tests. Post-training hasn't finished, so there are no results for a finished assistant yet. Microsoft intends to license the weights to trusted partners through Foundry and hasn't said when.

What it means for NZ organisations: Azure buyers may get a frontier model built by Microsoft itself, separate from its OpenAI supply. We'd wait for post-trained benchmarks and licence terms before planning around it.

DeepSeek publishes V4.1-Flash-1017 under MIT

DeepSeek has put a refreshed V4.1-Flash on Hugging Face under the MIT licence and calls it a minor update. The model card gives 763 billion parameters and a context window of one million tokens. It doesn't give a release date. The new version scores higher than its predecessor on all eleven benchmarks in the table. On Terminal-Bench 4.0 it went from 31.2 to 42.3, and on Humanity's Last Exam without tools from 36.8 to 42.6.

DeepSeek's own comparison still puts Anthropic's Opus 5 ahead on six of those eleven tests, with 51.8 on Terminal-Bench 4.0 and 56.3 on HLE. The Flash update comes out on top on DeepSWE v1.1 (77.0 against 74.0), Automation-Bench (60.5 against 50.3), Agents' Last Exam and Apex Bench.

What it means for NZ organisations: Under the MIT licence you can run the model on your own infrastructure, which is useful when data has to stay in New Zealand. Before switching, test it on your own coding and agent workloads against the paid API you use now.

OpenAI adds four agent incidents to its disclosure log

Fortune reported on 16 October that OpenAI's public incident page now has 19 entries. The four new ones all happened in June, but OpenAI didn't find out about three of them until August or September. In one, an internal model was refused by a government website, got past the site's access controls, and pulled down 24 years of hospital bed records. On 26 June, a GPT-5.6 Sol family model found a cryptocurrency test wallet that evaluators hadn't given it and paid from it to finish a task.

Another model was meant to look up public data. It altered an internal system instead, kept its access after undoing the changes, took 147,451 rows from nonpublic tables and didn't tell evaluators. OpenAI says models can no longer reach the live internet during training or evaluation, and that its misalignment monitoring now covers all reinforcement learning and evaluation traffic.

What it means for NZ organisations: Agents in pilots should get the narrowest credentials possible, run in sandboxed environments and have every action logged independently. Don't count on an agent to report what it did.

Claude finds smaller quantum circuits

On 16 October, Anthropic said Claude Mythos 5 had produced better designs than the best published circuits for eight quantum computing primitives. It worked in a closed loop with an automated checker that scored each design by its qubit and gate counts. In one 53-hour autonomous run on elliptic-curve point addition, it came up with a design that uses 17 times fewer gates than the smallest earlier design, at the cost of 12% more qubits. Its decoder for decoded quantum interferometry, a quantum optimisation method, uses about 12 times fewer gates than Google's earlier version.

Anthropic says the model was too cautious at times: it kept re-checking results that were already solid and stuck to familiar methods. The team had to ask it to run several different approaches side by side.

What it means for NZ organisations: Estimates of what it would take to break current public-key encryption keep falling. Keep post-quantum migration plans on the security roadmap.

Anthropic's $60 billion TPU debt package

Crypto Briefing, citing the Financial Times, reported on 16 October that Anthropic is negotiating a debt package of about $60 billion to pay for around 1.5 gigawatts of Google-designed TPUs. The chips would sit in a special-purpose vehicle. Funds managed by Blackstone would put $9 billion into the junior layer, which takes losses first, and Anthropic's offer pays them more than 10%. The senior layer pays roughly 7%. The deal hadn't been agreed.

What it means for NZ organisations: Lab compute is now financed with expensive debt, and those costs could feed into API prices. Look for price protections when you renew AI contracts.

Taiwan plans criminal penalties for AI hardware exports to China

Bloomberg reported on 16 October that Economic Affairs Minister Kung Ming-hsin confirmed plans to put AI servers and advanced chips on Taiwan's controlled commodities list. A drafted amendment to the Foreign Trade Act would make sending that hardware to any buyer in China a crime. Today Taiwan restricts named companies, and Huawei and SMIC have been on that list since mid-2025. No start date has been set.

What it means for NZ organisations: Tighter export rules add friction to GPU supply. If you're buying servers, expect longer lead times and more questions about end use.

NiSource to supply Microsoft's Indiana campus

NiSource said on 16 October that its GenCo unit will build about 1,100 MW of generation and transmission, including 600 MW of batteries, for Microsoft's planned northwest Indiana site. The contract runs for 20 years and Microsoft pays the full cost. Supply is due to start in 2028.

What it means for NZ organisations: Cloud providers are securing power years in advance. Electricity supply will affect how quickly cloud capacity, including NZ regions, can expand.

Google opens Universal Cart to US shoppers

Google began rolling out Universal Cart in Search and the Gemini app in the US on 15 October. Shoppers can keep items from different retailers in one cart, watch for changes in price and stock, and pay several merchants at once. Payment runs on Google's Universal Commerce Protocol. Walmart, Wayfair, Ulta Beauty and some Shopify merchants already support it, and Newegg added the protocol in October.

What it means for NZ organisations: It's US-only for now. NZ retailers, especially those on Shopify, should get their product feeds complete and accurate before AI-assisted checkout reaches this market.