Afternoon edition — Monday 19 October 2026

MiniMax said on 19 October that it is rolling out M3.1 Flash through the MiniMax API, its M Plan subscription and its MiniMax Code app. The company said Space Bunny, a model that processed more than 200 trillion tokens while operating anonymously, was an early version of M3.1 Flash. MiniMax said model weights will be released later, and that a larger Pro model will also arrive later.

MiniMax lists flat pricing of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. It claims online inference is nearly three times faster and agentic-task usage costs are 70% lower than its predecessor, with output speeds above 200 tokens per second and a median above 170.

In its own benchmark results, MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, 76.2 on BankerToolBench and 82% on CyberGym.

MarkTechPost reports that MiniMax's product page lists 428 billion total parameters with about 23 billion active, a one-million-token context window and reasoning that cannot be switched off. It says the weights are due on 23 October under a license not yet disclosed, and that MiniMax's comparison table puts the model at 38.1 on Terminal Bench 4.0.

Amazon announced TorchNeuron, a native PyTorch backend for its AWS Trainium AI chips. Amazon says PyTorch code that developers already write runs on Trainium with a single configuration change and no rewrite, and that in a side-by-side demonstration a training script learned at the same rate on a conventional chip and on Trainium.

A post on the PyTorch Foundation blog by AWS engineers says TorchNeuron plugs into PyTorch's PrivateUse1 device backend and torch.accelerator APIs. It supports eager execution, torch.compile and the Neuron Kernel Interface for custom kernels on Trainium2 and Trainium3. The post says the code will be released under the Apache 2.0 license at aws-neuron/torch-neuronx as a public beta on 20 October, at PyTorch Conference North America.

Amazon says it is contributing changes upstream to make PyTorch less tied to specific chips and tests TorchNeuron against hundreds of thousands of models on every build. It names Cartesia and the Lieber Institute for Brain Development among early users.

NorForge published an investigation on 19 October into publicly exposed control panels for Cairn, an open-source AI agent platform it describes as a state-based problem-solving engine. An operator sets a starting point and a goal, and agent workers move toward it, posting discovered facts and planned next steps to a shared board.

NorForge says it found more than 20 exposed Cairn panels it assesses as likely malicious, all apparently financially motivated, with no obvious state-sponsored activity. Its largest cluster, which it labels AIA0001, uses Chinese, spans 14 panels and more than 1,100 attack projects, and mostly targets organizations in the United States, Japan and Australia, apparently seeking persistent access to e-commerce and payment infrastructure. Some exposed panels appeared to belong to legitimate red teams, which NorForge says could put their clients at risk.

The Verge, citing Reuters, reports that victims included John Deere, SharkNinja and Cummins, that the most active attacker broke into more than 100 websites, and that successful breaches took about 23 hours. John Deere told Reuters there was 'no indication that sensitive information was accessed', and SharkNinja said a standalone product-support site was affected and the threat contained. The Verge reports that Cairn's creator said earlier this month that the platform 'was never a hacking tool'.

The US Department of Homeland Security filed a notice for public inspection on 19 October, due to be published in the Federal Register on 20 October. It says the Secretary of Homeland Security has ordered two things: information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd, including affiliates, are excluded from federal civilian agency procurements; and all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations, are to be removed from civilian agency IT.

The notice cites the Federal Acquisition Supply Chain Security Act of 2018 and a review of a Federal Acquisition Security Council recommendation. The signed order is filed under docket DHS-2026-0628, and a CISA team is named as the implementation contact.

Reuters, as published by CNA, reports that DHS said DeepSeek models pose 'multiple threat vectors' and that their use 'may result in the transfer of sensitive data to the PRC if used by U.S. government employees or vendors.'

Tenstorrent said on 19 October that it has closed a $1.5 billion Series E round at a $6.2 billion post-money valuation. Trace Capital and Innovation X led the round, with participants including ARK Invest, Fidelity Management and Research Company, Morgan Stanley, Sandisk, T. Rowe Price, Tether Investments and XTX Markets.

The company says it will use the money to deploy its Galaxy servers at scale, keep serving open-weight language and video-generation models, and develop the AI and CPU designs it licenses to chipmakers. It names Samsung and LG as licensees of its TT-Ascalon RISC-V CPU and Tensix AI cores. Chief executive Jim Keller said: 'We built one architecture and sell it at every size.'