AI Digest Nov 16, 2026

AI progress — 16 November 2026

The most consequential story today is political: The Verge reports that Congress wants to tighten export rules on AI chips to China after a spying scandal involving Nvidia, which has also put Jensen Huang’s position in the news. Any tightening would change which hardware buyers in China can get and what Nvidia can ship there. Elsewhere, llama.cpp reached v1.0.0, Qwen published Qwen4-Max, and AI-assisted reading of Gödel’s notebooks drew attention. Details on the Nvidia scandal come from secondary reporting only.

New models and papers

  • Qwen4-Max — Qwen published a blog post announcing Qwen4-Max; the candidate feed carries only the title, so read the post for specifics. Why it matters: a new flagship from a major model family is worth checking against whatever you currently use.
  • MiniMax-H3-Image — an image model from MiniMax, published on Hugging Face and already available in Comfy-Org and GGUF community conversions. Why it matters: there are packaged versions for ComfyUI and GGUF runtimes, so it can be tried locally without custom conversion work.
  • Loopback Transformer: Towards a Unified View of Residual Looped Transformers — a paper that, per its title, proposes a unified view of residual looped transformers. Why it matters: for anyone following parameter-sharing architectures; check the paper for the authors’ reported results.

New tools and software

  • llama.cpp v1.0.0 — the local-inference project has cut a 1.0.0 release, described in a Hugging Face post as a stronger foundation for the local AI ecosystem. Why it matters: a 1.0 tag is the project’s own signal of a stable baseline to build local tooling on.
  • Gemini 4 Hydrogen in GitHub Copilot — GitHub added the Gemini 4 Hydrogen model to Copilot. Why it matters: Copilot users can select it now; a separate GitHub notice lists Copilot model deprecations coming in December.
  • Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo Switchyard — an AWS post on routing coding-agent requests between models on Bedrock using NeMo Switchyard. Why it matters: it is a documented way to split agent traffic across models instead of hard-coding one.

Overall digest

Share
Subscribe

Get the AI digest by email

One email per weekday morning: new models and papers, new tools, and the overall picture. No ads, unsubscribe in one click. See the privacy policy.