August 11, 2026
The agent model that fits on one GPU
Issue #1 (Tuesday Debrief) - Aug 11, 2026, 10:00 AM ET. Audience = warmup-batch-1 (290 never-mailed) + website tag (174). Staged send: remainder goes Friday if bounce holds under 3%. Drew reviews and schedules.
Let's debrief.
Meta just put a genuinely capable agent model on hardware you already own, and the license is the part almost nobody is talking about.
📌 The big thing
Meta open-sourced a 30B agent model that runs on one consumer GPU
Meta released Muse Glimmer on Monday: a 30-billion-parameter agentic model shipped under Apache 2.0. A 30B model normally needs over 55 GB of memory at full precision. Meta compressed it to roughly 4-bit, which drops it under 20 GB and fits it inside a single 24 GB or 32 GB card. On an RTX 5090 it decodes at 233 tokens per second. On an M4 Max, 38. Weights and quantized builds are on Hugging Face today.
The take: the license is the story, not the benchmark. Llama always shipped with usage restrictions attached. Apache 2.0 does not. Pair a permissive license with a model small enough to run on your own machine and the marginal cost of an always-on agent goes to roughly zero. When an agent costs nothing per call, you stop designing around how many calls you can afford and start designing around what you actually want it to do. That is a different kind of product.
Why you care: if you shelved a background agent because the API bill scaled with usage, that constraint just moved. One caveat before you build on it. Meta's own numbers put Glimmer ahead on tool orchestration and behind Qwen3.6-27B on computer-use and terminal work, and those numbers have not been independently replicated yet. Treat it as a tool-calling brain, not a click-the-screen agent.
⚡ Worth your time
→ The EU AI Act deadline everyone reported on August 2 was not the one that actually landed. High-risk obligations got pushed to December 2027. What turned on instead was Article 50 transparency: if you ship a chatbot to EU users it has to identify itself as AI, and synthetic media has to be labeled. That part is live right now. What changed →
→ OpenAI split its Daybreak security program into two tiers and put GPT-5.6-Cyber behind the vetted one. In testing it found two unknown bugs in Chrome's V8 engine that chain into a sandbox bypass, since patched. Notice the shape of it: the capability is here, and the thing gating it is an application form. The details →
→ Full-duplex voice quietly got good this week. NVIDIA put out an open speech-to-speech model with roughly 450 ms turn-taking and live tool calling. If you tried building a voice agent a year ago and gave up on the lag, it is worth another hour of your time.
🙅 Skip it
The "Google's AI will call stores for you" headlines going around again. It is real, but it was announced last November and is still finishing its US rollout. Nothing actually changed this week. If you sell anything locally it is worth preparing for eventually, just not today.
The tools and the people behind this issue live in AI Academy. It is free, and it is where I answer things I cannot fit in an email. Come hang out →
That's the Debrief.
See you Friday for the Toolkit.
Talk soon,
Drew
P.S. What are you building right now? Hit reply and tell me. I read every one.