A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 6 STORIES

AI news, 16 November 2026

From the community

Alibaba's Qwen4-Max is out, and cheaper than the last version

Alibaba has fully launched Qwen4-Max, its top model, after last week's preview. It can read images, handle very long documents (up to a million tokens) and is available through Alibaba's QwenCloud API, Qwen Studio, QwenWork and Qoder, as well as OpenRouter and Vercel's AI Gateway. Prices are $1.50 per million input tokens and $4.50 per million output tokens, lower than Qwen3.8-Max. Alibaba showed it on 3D modelling, game building, turning sketches into manufacturing plans, finance, law, healthcare and science work.

Why it mattersIf you pay for AI by usage, this is a strong model at a lower price. Alibaba says it uses up to about half the tokens and cuts the average cost per task by over 60%, but those are the company's own figures from its own tests, so check on your own work before switching.

Sourcesx.com post ↗

Tencent's Hy4 flash is a very cheap model for everyday work

Tencent has released Hy4 flash, the budget version of its Hy4 model, aimed at everyday productivity tasks. Through its API it costs $0.070 per million input tokens and $0.281 per million output tokens, with cached input at $0.003.

Why it mattersFor high-volume routine jobs like drafting, summarising and sorting, price per task matters most. Tencent says Hy4 flash matched Claude Haiku 5.5 on daily work in blind tests at 18 times lower cost; that is Tencent's own claim and has not been independently checked.

Sourcesx.com post ↗

Samsung's LittleBit squeezes a 13B model into under 1 GB, but it's research code

@dr_cintas highlighted LittleBit, Samsung's open-source method for compressing an AI model's weights to below one bit each (down to 0.1 bits), which he says fits a 13-billion-parameter model in under 1 GB. It's real: Samsung Labs' official code for LittleBit (NeurIPS 2025) and LittleBit-2 (ICML 2026) is on GitHub, and Samsung Research reports Llama 2 13B at about 0.84 GB. But answers get worse than the full model's, it needs extra training and special software to run fast, the 'up to 11.6x faster' figure is the researchers' own, the code dates from November 2025, and the licence rules out commercial use. Our verdict: partly fact.

Why it mattersThis is where cheap, private AI on laptops and phones is heading, but it isn't something to deploy today. If you want local AI now, stick with standard compressed models that run in tools like Ollama, and watch this space for the next year.

Sourcesx.com post ↗ · github.com ↗Partly fact Hype or Fact? check

38% of new arXiv papers read as AI-written, by one detector's count

A scan of 13,500 papers on arXiv, the open research archive, found that 38% of those sampled from August to October 2026 read as machine-written, against a threshold set so that only 0.4% of pre-ChatGPT papers were flagged. Computer science was highest at 62% and maths lowest at 3.7%. The author is upfront about the limits: the detector measures writing that reads as AI, not proof of who wrote it, so it can't tell a fully generated paper from one an AI only edited, and maths may score low simply because there's less prose to judge. He also sells an AI 'humaniser', so this is one tool's estimate rather than an independent study. Shared by @kimmonismus. Our verdict: partly fact.

Why it mattersExpect clients, editors and AI tools to judge whether your reports and articles sound machine-written, so edit AI drafts into your own voice. And never use a detector score on its own to accuse a job candidate, student or supplier.

Sourcesx.com post ↗ · unslop.run ↗Partly fact Hype or Fact? check

json-render comes to Swift: AI-built screens drawn natively on iPhone and Mac

@ctatedev announced json-render for Swift. json-render, from Vercel Labs and in our directory, lets an AI describe a screen as structured data using only components you've approved, and your app draws it. The Swift version draws those screens with Apple's own SwiftUI on iOS and macOS, reads the same descriptions as the existing React version for the web, and supports your own components, saved state, actions and rules for what to show. When I checked on 16 November, the Swift package wasn't yet in the public json-render code repository, so treat it as just announced.

Why it mattersIf you have an app on the web and on Apple devices, one AI-generated layout could drive both, and because the AI can only use components you've designed, your brand and safety rules stay in your hands.

Sourcesx.com post ↗ · json-render.dev ↗

Facade Code runs several coding agents side by side in VS Code

Facade Code is a free, open-source (MIT) add-on for the VS Code editor that runs Claude Code, Codex and other coding agents side by side. One agent can start another, give it a task, message it or help it when it's stuck, and Git worktrees keep each agent's changes separate. It's young (about 390 stars on GitHub) and reached me through a promotional repost by @RealMarvelX, so try it on a test project first.

Why it mattersRunning several agents at once is how small teams get more done per developer, but only when each agent has a separate task you can review. Start with two agents, not ten.

Sourcesx.com post ↗ · github.com ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.