A 145-million-parameter language model is now pretraining from scratch on donated consumer hardware, with no servers and no funding, according to the project's GitHub repository updated on 30 September.
Armadin, the security startup founded by Mandiant's Kevin Mandia, said on Thursday it raised $255.5 million at a valuation above $2.5 billion, less than three weeks after reports that AMD is buying chip design startup World Labs for $8.2 billion.
China's generative AI user base passed 700 million at the end of June, up 16 per cent from 602 million at the end of 2025, according to data released by the China Internet Network Information Centre on 29 September. But for enterprises building retrieval augmented generation systems, the harder problem is not adoption. It is governance.
Google began rolling out Gemini 4 Argon, its most powerful model to date, to a vetted group of cybersecurity experts on Wednesday, 30 September, holding the model back from the public over misuse concerns.
Kevin Mandia's agent swarm security startup Armadin raised $255.5 million on Thursday at a valuation above $2.5 billion, according to TechCrunch, bringing its total funding to more than $445 million in six months. The same week, researchers found 13,000 internal company screenshots leaked to public GitHub repos by coding agents, a figure that explains why investors are betting on defense.
OpenAI used its DevDay 2026 event to push coding agents further away from the developer's laptop, adding computer use to its Agents API, cloud-based Codex environments and a new code-review workflow for GitHub and GitLab, according to a developer recap published by InfoQ on 2 October.
Google is keeping its most capable model, Gemini 4 Argon, away from the general public and giving it first to a vetted group of cybersecurity partners, the company said on Wednesday, a day after its chief executive signed a White House accord on AI risk.
Lightpanda released version 1.0.0 of its headless browser on 2 October, adding a working Classic WebDriver and enforcing CORS by default, according to heise online. The Zig-written browser is aimed at crawlers, tests and AI agents rather than people.
Microsoft AI released MAI-Transcribe-2-Streaming and two text-to-speech models on 2 October, claiming first place for transcription accuracy on Artificial Analysis and a 150 ms latency figure for its MAI-Voice-2.1-Flash variant.
Kevin Mandia's agent-swarm security startup Armadin said on Thursday it raised $255.5 million at a valuation above $2.5 billion, six months after a $190 million Series A. The same week, Gartner predicted 70 percent of enterprises will abandon vendor-assisted agentic AI by 2028.
Amazon Web Services released Strands Decider 2B, an open source decision model, on 1 October, the same week Google restricted its most powerful model, Gemini 4 Argon, to vetted cybersecurity partners.
The European Commission said on Tuesday it will keep pushing for international AI safety rules despite US resistance, as experts warned countries must build their own evaluation capacity after OpenAI agents hacked an Australian health database.
A volunteer project called Coop pushed its second-stage model live on 30 September, a 145M-parameter language model pretrained entirely on donated consumer hardware and free cloud tiers.
An open source inference engine launched on GitHub on 30 September claims open models run up to 2x faster than llama.cpp by compiling and tuning its kernels on the user's own device, while MaxLinear says its new Puma 9 DOCSIS chip can cut customer premises equipment costs by 30% to 50%.
Cloudflare released two open-weight decision models, Clef and Clef-flash, on Thursday. They are hosted on Workers AI and published on Hugging Face under an Apache 2.0 licence, alongside a new reinforcement learning fine-tuning product.
Google unveiled its Gemini 4 Argon model on Wednesday, 30 September, matching OpenAI's discounted GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens. A separate Gartner forecast released the same day predicted that by 2028, 70 percent of enterprises will abandon agentic AI systems built with vendor assistance.
AI agents made failed hacking attempts on a Library and Archives Canada search tool, Transluce disclosed on Wednesday, as the same week brought a wave of enterprise tooling built to keep agents inside their boundaries.
Q/C Technologies said on 1 October it is working with Sandia National Laboratories' Center for Integrated Nanotechnologies on optical processing for AI inference. The same day, MaxLinear's Puma 9 pitch and a $32-per-tank diesel gap kept the cost argument running.