AI·Digest
← All digests
evening

AI Digest — July 16, 2026

🧠 Models & Releases🛠️ Tooling & Dev🏢 Industry & Business⚖️ Policy & Safety

Quick Notes

  • NVIDIA Nemotron 3 Embed — NVIDIA released a collection of open, commercially available embedding models for RAG and agentic retrieval; the 8B BF16 variant ranks #1 on the RTEB multilingual leaderboard, and 1B BF16/NVFP4 variants target production-scale throughput with NVFP4 acceleration on Blackwell. https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb

  • Kimi K3 — Moonshot AI announced Kimi K3, a 2.8T-parameter “open 3T-class” model (open weights promised by July 27) priced at $3/$15 per million tokens — its most expensive model yet — beating Claude Opus 4.8 and GPT-5.5 on self-reported benchmarks and topping Arena.ai’s Frontend Code arena. https://simonwillison.net/2026/Jul/16/kimi-k3/#atom-everything

  • Herdr terminal multiplexer — Herdr is an open-source, ~10MB Rust “agent-forward” terminal multiplexer (tmux-like) that runs multiple coding agents in one window, auto-detects 15+ agents, shows which need attention, and keeps agents running as a background server even after the window closes. https://chaseai.io/blog/herdr-terminal-multiplexer-ai-coding-agents

  • [Video] Herdr walkthrough — Chase AI demos Herdr running two Claude Code terminals plus OpenCode and Codex in one space, highlighting its three benefits: organization, multi-agent monitoring, and persistence via a background server. https://www.youtube.com/watch?v=neK8ydl0Vlk

Structured Summaries

New Models / Research

Three notable model releases landed today. NVIDIA Nemotron 3 Embed is a family of open, commercially usable embedding models built for production RAG, agentic retrieval, code retrieval, and agent memory; the 8B BF16 model ranks #1 on the RTEB multilingual leaderboard and delivers both the highest retrieval accuracy and the lowest downstream agentic token cost across ViDoRe V3, BRIGHT, and BrowseComp-Plus, while the 1B NVFP4 variant uses native NVFP4 acceleration and Quantization-Aware Distillation for high-throughput, lower-cost serving (shipping with an optimized NIM microservice). Kimi K3 from Moonshot AI is a 2.8-trillion-parameter model billed as the first “open 3T-class” model (open weights due July 27), taking the size crown from DeepSeek v4 Pro; it mostly beats Claude Opus 4.8 and GPT-5.5 on self-reported benchmarks, leads Arena.ai’s Frontend Code arena, and is priced at a Sonnet-level $3/$15 per million tokens — the priciest Chinese model to date. Inkling, from Mira Murati’s Thinking Machines Lab, is that lab’s first open-weights release: an Apache-2.0 MoE transformer with 975B total / 41B active parameters, trained on 45T tokens of text, image, audio, and video. It is explicitly not a frontier model but a customizable multimodal base for fine-tuning on the lab’s Tinker platform, with a smaller 276B (12B active) Inkling-Small still to come. Sources: https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb https://simonwillison.net/2026/Jul/16/kimi-k3/#atom-everything https://simonwillison.net/2026/Jul/16/inkling/#atom-everything

Product Launches

Herdr is an open-source, ~10MB single Rust binary that acts as an “agent-forward” terminal multiplexer for AI coding agents — essentially tmux, but with a dedicated agents tab that surfaces which agents are working, stuck, or waiting on input. It organizes work in a spaces/tabs/panes hierarchy, auto-detects 15+ agents (Claude Code, Codex, Copilot CLI, OpenCode, etc.) with zero config, and runs as a background server so closing a window doesn’t kill in-progress agent work. A companion Herdr skill lets agents spawn their own panes and spaces, and there’s a plugin marketplace forming around it. The Chase AI writeup and its companion video both demo running two Claude Code terminals plus OpenCode and Codex in one workspace, using Herdr to make a normally-headless Claude↔Codex plan-critique-revise loop visible in real time. Sources: https://chaseai.io/blog/herdr-terminal-multiplexer-ai-coding-agents https://www.youtube.com/watch?v=neK8ydl0Vlk

Research Insights

Firefox in WebAssembly is a striking demonstration of what current coding agents can accomplish: Puter compiled Firefox/Gecko (chosen for its strong single-process support) to WebAssembly so the entire browser runs inside another browser, an effort estimated at ~$25,000 of Claude Opus and Fable tokens on a Claude Max plan. Because in-browser code can’t open arbitrary network connections, all traffic is funneled over a WebSocket using the Wisp protocol through Puter’s servers (which had to scale up during the Hacker News surge); end-to-end encryption to HTTPS sites was confirmed in the WebSocket traffic. Separately, Simon Willison reflects on his 21-month-old “SVG of a pelican riding a bicycle” benchmark, concluding its once-surprising correlation with real model quality has now largely broken down (GLM-5.2 pelicans outclass GPT-5.6 and Claude Fable 5) — chiefly because it says nothing about the agentic tool-calling reliability that matters most today, though he still values it as a forcing function to actually run each new model. Sources: https://simonwillison.net/2026/Jul/16/firefox-in-webassembly/#atom-everything https://simonwillison.net/2026/Jul/16/kimi-k3/#atom-everything

Tooling

Two safety/tooling notes for coding agents. Pragmatic Engineer’s The Pulse reports that Grok’s CLI — gaining traction on the back of the capable Grok 4.5 coding model — was caught uploading all of a user’s local files to the cloud, a serious trust issue for a developer tool. In a similar vein, a quoted note from Thibault Sottiaux details a Codex bug in which GPT-5.6 unexpectedly deleted files under certain conditions. On the lighter side, Simon Willison’s mermaid-ascii experiment compiled the Go library AlexanderGrooff/mermaid-ascii to WebAssembly (via Claude Fable 5) to compare against his earlier Grok Build Rust-based Mermaid-to-ASCII converter. Sources: https://newsletter.pragmaticengineer.com/p/the-pulse-groks-cli-caught-uploading https://simonwillison.net/2026/Jul/16/bad-codex-bug/#atom-everything https://simonwillison.net/2026/Jul/16/mermaid-ascii/#atom-everything

🔗 View this digest on the web: https://ainews.rusig.com/digests/2026-07-16-evening/

#ai#digest