The Stack Report: The Missing MLX Port Arrived, And The Deciders Got Small

Five days late, but it arrived

Last week I told you EmbeddingGemma 2 could put your words and your photos into the same search space, and then I said I had seen nothing showing it ran on MLX. Six days later, MLX-VLM 0.7.8 shipped with EmbeddingGemma 2 multimodal embedding support built in.

It is Wednesday, 14 October, and a week has passed since the last one.

This matters more than a normal port because of where it landed. Imagebook already uses MLX-VLM. Since moondream left, that library has written your photo captions. So the door from last week no longer needs a new runtime. The model that captions your photos and the model that could embed them now come from the same package. No Transformers stack, no Ollama dependency for this lane, and one fewer moving part than last week's plan.

The idea has not changed. Today Imagebook search goes through bge-m3, so it reaches a photo only through the words in its caption. If a caption is wrong, search is wrong in the same way. Embedding the image directly removes that middle step. The short vector sizes keep the encrypted index small.

What has not changed is the test, and nobody has run it yet. Embed one album, run ten searches whose answers you already know, and count the hits. Do the same with today's caption-based search and compare. Like every model, the weights wait for your yes on the download. Plan for an afternoon, and most of it goes to choosing the ten searches.

The deciders got small

The same release added three more decision models to that runtime: something called Clef, and two from Liquid AI named d1-3B and d1-omni-600M. Last week Decider was one interface with one model behind it. Now the local side has a field, and its smallest entry is 600 million parameters. At that size it barely registers on your Mac.

This changes the Cohort D comparison I described last time. Capture spent this week on the card rounds and locked all five legs. Route proposal is still the next reshape step, and it still has not picked a mechanism. The replay I suggested, a few hundred captures you already routed by hand, now gets cheaper in both directions. The local models cost nothing per call, and the smallest is fast enough to run the whole set many times. Run the 600M and 3B models next to Luna's endpoint, with Mistral as the control. If the tiny one ties, you have a router with no API, no meter and no outage, and that is the kind of moving part you like most.

One footnote that advances last week's story. Codex 0.162 added a comparison path that sends its own Guardian approval reviews through the Decisions interface. OpenAI's own agent now uses the beta I told you to keep away from live traffic. That makes the interface more credible. It does not remove the beta label, so keep the rule.

Heads up

Haiku 5.5 launched on the 7th, and Ask already runs narration on it, so you are ahead on that one. The trap is everything else still on Haiku 4.5. On 5.5, sampling settings such as temperature, manual thinking budgets and assistant prefill each return an error. Adaptive thinking is also on by default, so responses spend tokens you did not ask for. Boxwatch picked Haiku in exp 097 because it was the cheapest model with saturated detection. If anyone moves that judge to 5.5, rerun the cost figure, because the thinking tokens change the answer the cost was based on.

FastAPI 0.143 reverses part of what I told you last week. Automatic OpenTelemetry exporter setup is now opt-in. Your canonical setup configures its own exporter over the popcorn2 socket, so most services are unaffected. The risk is any service that started relying on the framework's automatic setup during the template ruling. On upgrade, its telemetry stops without an error message. That is one of the fail-green classes Director already has a lesson for. A one-line grep for that setting before anything moves past 0.142 settles it.

uv 0.13 and Python 3.15.0 both landed. uv now downloads 3.15 when nothing is pinned. On your Mac it prefers interpreters you already have, so the risk is low there. Where it does bite is MLX projects. Ordagrant capped itself below 3.14 for a reason, and the other MLX repos should state the same cap rather than rely on what happens to be installed. uv is also stricter about tar archives, and it now enforces hash requirements in included constraints files. Neither affects your deploys as far as I can see.

Codex changed its default model to GPT-6.1 Sol in 0.161, during the same week Director froze an 18-branch trial set for exp 092 arm C. If codexburn branches straddle that switch, the model is a covariate the trial records should name explicitly.

The rest, briefly

FastMCP 4.1 is the line PärMCP's version 4 evaluation will land on. HTTP sessions now expire after 30 idle minutes, and MultiAuth client identities are qualified by source, so treat it as a migration. Stay on 3.4.8 for now.

Sonnet 5.5 cache reads now cost half as much. Ask sends its schema as a cached system prompt, so you save money without doing anything.

Max plans now include monthly API credits. Director is already tracking the second grant, which expires 13 November, so you are on that one already.

Managed Agents can now run dynamic workflows server-side. You paused stack-watch on Sunday, so this is for later, if at all.

R2 now breaks bandwidth down by Cloudflare location. That is a third way, after last time's two, to see who besides you downloads the agent feeds.

The Podman releases patch gvproxy, which is used by the Mac and WSL machine layer. Popcorn2 does not use gvproxy, so this one is not for you. Scaleway's new Milan region and private serverless endpoints are not for you either. Neither are Astro's two patch releases or OpenAI's Ultrafast tier.

MLX-VLM also added MiniCPM-V 4.7 in a 1-billion-parameter version, one more small captioner to try if Qwen ever disappoints knarfbot.

If you look at one thing

Embed one Imagebook album with EmbeddingGemma 2 on MLX and run your ten known-answer searches. The model now sits in the library you already run.