The Stack Report: A Thirty Billion Parameter Agent That Now Runs On Your Own Mac

The model that showed up on Apple Silicon

Meta Superintelligence Labs released its first open model this week, and Ollama shipped support for it on Apple Silicon the same day. It is called Muse Glimmer. Thirty billion parameters, multimodal, and built specifically for agent workloads rather than chat. It takes images. It runs through Ollama's MLX engine, which is your local default anyway. And the part that matters: Ollama ships a launcher that points Claude Code, Codex, or a long-running assistant framework at it instead of at a paid endpoint.

Two independent channels confirmed it in the same week. The MLX vision-language library added Muse Glimmer support and a tool-call parser for it in the same release. That is usually the signal that something is real rather than announced.

Here is the door, and it is a specific one. Autocast gates every scheduled show on your Claude subscription meter, because the separate headless budget got paused back in June and never came back. When the weekly meter crosses the cap, shows defer. One of them has been sitting on a stranded SMS ask since July because of exactly that. A local agent-shaped model with tool calling does not solve the creative half of that pipeline, and you should not point The Director Report at it. But the mechanical lanes underneath it, the source gathering, the relevance filtering, the classification steps that currently burn subscription headroom for no editorial gain, those are the shape this model was built for.

What it would take: an afternoon to pull it down and see whether it can actually hold a tool loop, and a second session to decide which autocast steps are dumb enough to delegate. Slow home internet applies. This is a thirty billion parameter download.

Dubbing that keeps the voice

ElevenLabs put Dubbing version two on the API. It translates audio into more than ninety languages and preserves each speaker's voice, tone and pacing. That alone is not new in the world, but the shape of the API is the interesting part.

It is project-based. Source transcripts and translations stay as editable JSON, you can edit individual segments, and then you regenerate only the regions that changed. That is the same in-place re-render pattern the PärPod editor already implements for annotations, except applied across a language boundary.

You have Swedish surfaces sitting right there. The private feeds you built for your sister already carry dual-language bodies and titles. Årebladet is Swedish-facing. And every automated show you run is English-only right now because that is what the pipeline does. Dubbing version two makes a Swedish track of an existing episode a rendering step rather than a rewrite.

The honest caveat: this bills per use, it is a new external dependency in a pipeline that already talks to Inworld and ElevenLabs, and nobody has listened to how a synthetic Graham sounds in Swedish. That is a render-and-listen question, not one to reason about.

A metered lane for the shows that keep deferring

Anthropic added session budgets to its Managed Agents product. A session gets a hard spend cap at public list rates, and when it reaches the cap it pauses with a specific stop reason instead of quietly continuing. Change the budget and it resumes. Sessions can also load skills straight out of a GitHub repository, and you can now pin which geography inference runs in.

The reason this is worth a paragraph rather than a shrug: it is a completely separate meter from your subscription. Your scheduled shows currently compete with your own interactive sessions for the same five-hour and weekly windows, which is the entire reason the SMS gate exists. A bounded, hard-capped, pay-per-run lane is the other answer to that problem, and it is the one that does not involve asking you for permission by text message at eighty-five percent.

The geography pin also lines up with the European-provider preference you hold everywhere else in the stack.

I would not move anything today. But when the next scheduled show starves for a week, this is now a real alternative to tuning the tiers again.

Codex learned to carry skills

Codex shipped portable agent plugins this week, with search across local, personal, workspace and remote catalogs. It also learned to import skills managed by Cursor, and to synchronise changes back to imported Claude and Cursor conversations without duplicating them.

Your whole working method is skill-shaped, and it runs across two models. This is the first release where the Codex side can actually carry the same portable bundles rather than being handed instructions in a prompt. Not urgent. But if the cowork and reverse-cowork lanes keep growing, this is the mechanism that stops them drifting apart.

Also in that release: an approve-for-me flag that runs approvals through an automatic review. Worth remembering the last time you tried to make a headless lane approve its own compound commands. That went badly enough to be written into the config as do not re-attempt. This is a different mechanism, so it is a maybe, not a fix.

Two things about to bite

Hugo removed Tailwind from its default security exec allow-list. If any site in the box-side build pipeline compiles Tailwind, the next Hugo bump fails at build time rather than at review time, and it fails on the server, not on your Mac. Check before upgrading the toolchain, not after.

Codex removed the full auto flag from its exec command entirely. Anything headless that still passes it now errors instead of degrading. The replacement is a sandbox flag. Worth one grep across the skills before the next unattended run.

Everything else was housekeeping. The uv package manager shipped two releases of performance work and a tool audit that is still behind a preview flag. Scaleway announced a service agreement for its ClickHouse warehouse and a Splunk integration, neither of which you own. Cloudflare turned on billing for R2 SQL, which you do not use, and which does not touch your podcast storage. And NVIDIA released a rather good-looking thirty billion parameter agent model that requires hardware you do not have.

One piece of good news with no work attached: the introductory price on Sonnet five is now the permanent price. The increase scheduled for September has been cancelled.

If you look at one thing

Pull Muse Glimmer down and find out whether it can hold a tool loop for more than three turns. That single answer decides whether the cheap half of autocast has somewhere else to live.