The Stack Report: A Second Local Agent Candidate, And An EU Lane For A Model You Have Not Chosen Yet

The one that arrived a week late

Last time I told you to pull Muse Glimmer down and find out whether it can hold a tool loop. That advice still stands, but a second candidate landed 2 days later and it is the one I would actually reach for first. Qwen 3.8, 27 billion parameters, and Ollama shipped a build specifically optimised for Apple Silicon with a stated purpose: repeated tasks and coding agents.

It is Wednesday, 19 August, and 7 days have passed since I last had anything worth saying.

The reason the smaller model is the better bet is the one I glossed over last week. Muse Glimmer is 30 billion parameters and multimodal, which means you pay for a vision tower that autocast will never use. Qwen 3.8 at 27 billion has a variant that exists purely because somebody sat down and tuned it for agents that run the same shape of task over and over. That is exactly what the mechanical half of autocast is. The source gathering, the relevance filtering, the classification steps that currently burn subscription headroom for no editorial gain.

The test is the same test. Can it hold a tool loop for more than 3 turns without losing the thread. If it can, the cheap lanes have somewhere else to live and your weekly meter stops being the thing that decides whether a show publishes. If it cannot, you have spent an afternoon and a large download, and you know.

A European home for a model you have been avoiding deciding about

Scaleway put DeepSeek V4-Flash on its Generative APIs, and this is the item I would act on soonest, because it costs you almost nothing and it closes something that has been open for months.

Your own editorial review default has been in a strange state since the DeepSeek alias silently started serving V4-Flash instead of V3. The lab notes say the proper successor pick is still to be decided, with V4-Flash as the leading candidate, pending a proper twelve-text multi-judge replication that nobody has run. So the default is V4-Flash by accident rather than by choice.

Scaleway hosting it turns that from a stalled decision into a cheap one. You already run on Scaleway. Your standing preference is European providers when quality is roughly equal. And the replication you never ran is not blocked by anything except somebody pointing the harness at an endpoint. If V4-Flash wins the bake-off, you get the pick you were probably going to make anyway, on an EU endpoint, with no new vendor relationship. If it loses, you have finally closed a question that has been marked open since June.

That is one session, and most of it is running an evaluation you already built.

The search stack grew a local reranker

The MLX vision-language library shipped a large release this week, and two things in it are worth your attention.

The first is reranking. It added both a text and multimodal reranking endpoint, and native sequence-classifier rerankers underneath it. Reranking is the step that sits between a search returning 50 plausible results and a human reading the 5 that actually answer the question. Your unified search across the mail archive, the newspaper archive and the conversation archive currently has no such step. It retrieves and it stops.

That matters more than it sounds, because you use those archives constantly and the top result is frequently not the right one. A local reranker over the top 30 hits costs nothing per query, runs on your own machine, and the whole integration is a single call inserted after retrieval. This is a small piece of plumbing with a disproportionate effect on a tool you touch every week.

The second is GOT-OCR 2.0, now supported in the same library. You have a newspaper OCR archive going back to 2019, and the quality of that archive is fixed at whatever the original OCR pass produced. A modern document model running locally means a re-OCR pass is now free in money and expensive only in patience. Whether that is worth doing depends on how often the archive search misses because of bad character recognition rather than bad ranking, and I would find that out before committing to a re-run of 7 years of scans.

Codex learned to tell you what a thread costs

Codex shipped one release, and it is a substantial one. The item that lands closest to how you actually work: estimated thread credits and cost now appear in the status command, in status lines, and in terminal titles.

Your entire lane choice runs off Codex headroom. Under 50 percent weekly means Codex implements and Opus reviews. Over 50 percent means the reverse. That threshold is currently checked once at the start of a session against an account-wide number, which tells you nothing about what any particular piece of work is costing while it happens. Per-thread visibility is the missing half of that. It will not change the rule, but it turns a start-of-session guess into something you can watch drift.

Two more from the same release. Hooks can now run asynchronously and can invoke tools over the model context protocol, which is a real capability change for a workflow with as many hooks as yours. And sessions can be forked from the command line, and archived or restored from the resume picker, which is the closest thing yet to an official version of the park-and-resume pattern you built by hand.

One thing that will bite

The MLX core library released 0.32.1 this week, and the vision-language library released 0.6.15 the same day, whose entire content is fixing itself against 0.32.1. Those two move together and they do not move alone.

Anything on this Mac that pins one and upgrades the other will break, and it will break at import time in a batch job rather than in front of you. The LoRA pipeline, the wardrobe identifier and the image tooling all sit on this pair. Upgrade them together or upgrade neither.

Everything else

The uv package manager shipped 2 releases of resolver performance work, some new CPython builds, and a preview flag that lets its check command install dependencies without building the project. Nothing that bites.

Cloudflare added a United States jurisdiction to R2 with guaranteed domestic data residency. Genuinely useful, and precisely the opposite of what you want. Not for you.

ElevenLabs shipped a large batch, and almost all of it is for the conversational agents product: queueing when an agent hits its concurrency limit, conversation summaries, topic pagination. Not your lane. The one piece that is nearly interesting is a new asynchronous generation API for image, video and speech, with reusable uploaded assets that let generation jobs be chained together. Chaining is the part with a future in it, since your sound pipeline is already a chain of prompts and cached clips. But it is a job-queue shape and your renderer is synchronous per segment, so this is a note for later, not work for now.

Anthropic renamed the Workbench to Playground. That is the entire entry.

If you look at one thing

Point your editorial evaluation at DeepSeek V4-Flash on Scaleway and finish the bake-off you paused in June. It is the shortest path from an open decision to a closed one, and the endpoint is on infrastructure you already pay for.