A Tuesday of open-model one-upmanship: DeepSeek's Flash update tops Claude Opus 5 on coding and MiniMax unmasks Space Bunny as M3.1 Flash. Off the wires: Washington orders DeepSeek out of federal systems, OpenAI races a $30B round past Anthropic's IPO, and Meta shows the supercomputer behind Muse.
The Department of Homeland Security ordered US civilian agencies to remove DeepSeek products, including third-party services built on its models, from their IT systems and barred new purchases. The order was signed July 24 and made public October 19; DHS cites 'multiple threat vectors,' warning that use could transfer sensitive data to China.
Reuters reports OpenAI is rushing to secure about $30 billion at a valuation above $1.4 trillion before Anthropic's expected mid-November IPO, with SoftBank and Abu Dhabi's MGX in talks to join. Separately, Bloomberg reports former Google CEO Eric Schmidt owns about 1% of Anthropic, bought in its Series A.
Meta published a detailed technical account of Prometheus, the gigawatt-scale AI supercluster in Ohio whose first phase trained the Muse models, and said it plans to scale the system to roughly one million GPUs. The post reports over 85% training goodput and previews Hyperion, its next supercluster in Louisiana.
Anthropic granted the nonprofit AI safety group METR 'employee-like' access to its systems for voluntary frontier-risk assessments, covering alignment, whether monitoring catches serious harm in time, and whether AI-accelerated research is outpacing safeguards.
DeepSeek released DeepSeek-V4.1-Flash-1017, calling it a minor update to September's V4.1-Flash with substantial improvements across reasoning, coding, agentic tasks and visual benchmarks, plus faster task completion and shorter outputs. Weights are open on Hugging Face under the MIT license. In DeepSeek's own tables, DeepSWE v1.1 rises to 77.0 from 74.2, ahead of Claude Opus 5 at 74.0; Terminal-Bench 4.0 jumps to 42.3 from 31.2; and all 11 reported benchmarks improve. API prices are unchanged, and DeepSeek also opened Search and Fetch APIs at $0.75 per 1,000 calls.
Every score here is DeepSeek's own, not independently verified, and Opus 5 still leads on HLE and Terminal-Bench. Even so, an MIT-licensed model beating a flagship closed model on agentic coding at a fraction of the serving cost keeps compressing the gap between open and closed AI.
MiniMax officially launched M3.1 Flash, a 428B-parameter mixture-of-experts model with about 23B active parameters, built for agentic tasks, coding, computer use and cybersecurity, and confirmed the viral anonymous model 'Space Bunny' was an early build. The company reports 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2 and 82% on CyberGym, with inference nearly three times faster and 70% lower cost than its predecessor at $0.15/$0.60 per million tokens. Model weights will be released later.
Two Chinese labs shipped flagship-class agent models within 72 hours of each other. When near-frontier agentic performance costs fifteen cents per million input tokens, ambitious multi-step agent workflows stop being a research luxury.
Editor's note: no new computational pathology papers in this morning's sweep. The journal scan's keyword hits were all off-topic, two arXiv runs found nothing in-window (the Monday announcement had not yet entered the API index at press time), and no lab announced new pathology research. For recent pathology AI, see No. 019: Beyond the Dominant Label, Evidence Field, and the chest X-ray foundation model benchmark.
Editor's note: no standalone agent papers beyond today's Selections. Both arXiv runs came back empty in-window, and the journal scan's reinforcement-learning hit was about network science, not agents. The agent story today is the two open model releases above: DeepSeek-V4.1-Flash-1017 and MiniMax M3.1 Flash.
Editor's note: no new fairness papers in this morning's sweep. The journal scan's keyword hits (a cholera vaccine study, a photoreactor paper) were off-topic, both arXiv runs found nothing in-window, and no lab announced fairness research. For recent work, see No. 020: cervical cancer screening disparities across 300 million patient records.
VAST open-sourced EdenLive, a real-time generative renderer that turns a reference video, from a game engine, a coding agent or recorded footage, into detailed video at 24 fps or higher on a consumer GPU. The release includes 5B and 14B diffusion-transformer weights, streaming inference code, and a prompt-to-playable-game pipeline, mostly under Apache 2.0.
World models have mostly been data-center demos. A real-time, open renderer that runs on a desktop GPU turns interactive worlds into something small teams can build with, from games to robot training data.