← archive

Get 5 AI picks like these every weekday at 7am MT. Free, plus 50 vibe-coding prompts when you confirm.

Post on XLinkedIn
Short Stack #5Tue, Oct 20, 2026
Image from inceptionlabs.ai: Mercury 3 is a diffusion LLM that writes at 1,000+ tokens a second
01

Mercury 3 is a diffusion LLM that writes at 1,000+ tokens a second

Inception launched Mercury 3 on Oct 19. It generates text by diffusion instead of word by word, and Inception says it runs at 1,068 tokens per second with quality on par with or better than GPT-6 Luna and Claude Haiku 5.5 (vendor claims). It has a 1M-token context, costs $0.10 in / $0.35 out per million tokens at a 50% launch discount, and is OpenAI-API compatible, on OpenRouter and Vercel too.

⏱ 10-MIN TRY

Take one OpenAI-SDK call in your app, point it at Mercury 3 using the free API tokens, and compare speed and output against your current model.

Source: inceptionlabs.ai

Image from minimax.io: MiniMax M3.1 Flash targets agent work at $0.15 per million input tokens
02

MiniMax M3.1 Flash targets agent work at $0.15 per million input tokens

MiniMax rolled out M3.1 Flash on Oct 19, priced at $0.15 in, $0.60 out and $0.015 cached input per million tokens. Its own numbers: 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 200+ tokens per second, and 70% lower agent-task cost than its predecessor. It's live on the MiniMax API and MiniMax Code; weights come later.

⏱ 10-MIN TRY

Rerun one agent task you've already logged on M3.1 Flash via platform.minimax.io and compare steps, cost and whether it finished.

Source: minimax.io

Illustration: a style curve above a row of color swatches
03

Lovart 2.0 adds Arti, a design agent that remembers your taste

Lovart 2.0 (Oct 19) adds Arti, which keeps editable memory of your aesthetic and working style. You can hand it whole jobs ("design delegation") on a schedule with a per-run credit limit and a test run first. A background Design Review flags off-style frames, and workflows you repeat 3+ times can become reusable skills.

⏱ 10-MIN TRY

Tell Arti two preferences (palette, how many directions you want first), start the weekly inspiration mood board template, and do a test run with a low credit limit.

Source: lovart.ai

Illustration: a flag with a prohibition sign
04

DHS orders DeepSeek out of federal civilian agencies

DHS made public on Monday an order, issued July 24, that requires federal civilian agencies to remove DeepSeek technology from their IT systems and stop buying it, CNBC reports. It covers third-party products and services built with DeepSeek's tech, and excludes the Defense Department. DHS cites possible data transfers to China and security flaws.

⏱ 10-MIN TRY

If you sell to (or hope to sell to) federal agencies, list every model and API behind your product, including routers and open weights, and flag any DeepSeek dependency.

Source: cnbc.com

Image from workspaceupdates.googleblog.com: Gemini apps now export to Google Sites in one click
05

Gemini apps now export to Google Sites in one click

Since Oct 19, any interactive app you build in the Gemini app can go to Google Sites via Export > Google Sites. It lands as a private, fully interactive draft you can edit, publish and share. Available now on web, Android and iOS for Workspace and personal accounts.

⏱ 10-MIN TRY

Ask Gemini for a tiny interactive tool (a pricing calculator or quiz), export it to Sites, add a heading, and send the link to one person.

Source: workspaceupdates.googleblog.com

Post on XLinkedIn

Get tomorrow's in your inbox.