US Homeland Security Orders Agencies to Remove DeepSeek

The US Department of Homeland Security (DHS) has told every civilian federal agency to remove products from DeepSeek, a Chinese AI company, and to stop buying them. Reuters reported the notice on October 19, 2026. The order is dated July. It also covers outside tools built on DeepSeek's models, so a federal employee's app can fall under it even if the DeepSeek name never appears. DHS gives two reasons: sensitive data could be passed to the Chinese government, and the models could contain malware or security holes. The Commerce Department, the Navy, New York, Tennessee and Virginia already had narrower limits. Reuters ↗

GitHub Copilot Switches Off Six AI Models

GitHub removed six models from Copilot, its coding assistant, on October 19, 2026: GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5. They are gone from every part of Copilot, including chat and autocomplete. GitHub's changelog names Gemini 3.8 Flash, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7 as replacements. If your company's Copilot lost a model and nothing new showed up, an admin has to turn the replacements on in Copilot's settings. GitHub first announced the cutoff on September 18. GitHub changelog ↗

Meta Details Muse Glimmer, a Model for Home Computers

Meta published a technical report on October 19, 2026 for Muse Glimmer, a free model built to do multi-step tasks on one home computer. Glimmer is mid-sized, with 30 billion parameters (the numbers that store what a model learned). Meta says a compressed version runs on a Mac or a PC with a single gaming graphics card. People using Ollama, a free app for running models at home, report an 18GB download. The license is Apache 2.0, which allows commercial use. In Meta's own tests, Glimmer scored 75.5 at using tools, ahead of Google's Gemma4-31B (54.2) and Alibaba's Qwen3.6-27B (62.5). Qwen still scored higher at operating a computer screen. Meta's report ↗

MiniMax Releases M3.1 Flash at Less Than Half the Cost

MiniMax, a Chinese AI lab, released M3.1 Flash on October 19, 2026. M3.1 Flash is a lower-cost model for AI agents, which carry out long tasks step by step. MiniMax says it was the anonymous "Space Bunny" model people had already been testing. Artificial Analysis, an independent testing firm, gave it 42 on its overall score, up from 29 for MiniMax's last model. Each test task cost about 22 cents, down from 51. One weak spot showed up. When M3.1 Flash did not know an answer, it guessed wrong 93% of the time and almost never said "I don't know." MiniMax blog ↗

Ant Group Releases Free Ming-3.0 Voice Models

Ant Group, the Chinese finance company behind Alipay, released five free Ming-3.0 speech models on October 19, 2026 under the MIT license, which allows commercial use. The smallest, Ming-3.0-tts, turns text into speech. It can create new voices or copy an existing one. It can also add sounds such as laughs and sighs. On Artificial Analysis's test of reading tricky words correctly, it scored 93.2%. That ties Fish Audio's Drama 3 Preview and beats ElevenLabs v4 (91.7%), a paid voice service many people use. Ant Ling announcement ↗

Claude Code Now Warns Before a Resume Uses Up Your Limit

Anthropic released Claude Code 2.1.304 on October 19, 2026. Claude Code is Anthropic's coding assistant, and it runs in a computer's terminal. The update adds a warning for Pro and Max subscribers before resuming an old session would use up their five-hour usage allowance. The /usage command also changed. It now shows when a rate limit ends. changelog ↗

scientific-agent-skills Adds Science Tools to AI Assistants

scientific-agent-skills gained 4,355 stars on GitHub this week. The free collection adds 177 ready-made skills to assistants such as Claude Code, Cursor and Codex. Each skill is a set of instructions for one research job. Most cover biology or chemistry, and medicine is another main area. The collection also links to more than 100 scientific databases, so a researcher can ask the assistant to look things up there. Read more ↗

ruflo Coordinates Teams of AI Coding Agents

ruflo gained 2,543 stars on GitHub this week. The free tool runs several AI agents at once and splits one large job among them. It works with Claude Code and Codex. It also lets the agents keep notes they can reuse in later sessions. Read more ↗