AI SYSTEMS

Five models on the bench.

Leaderboards rank models in general, which is a question nobody's workflow actually asks. A bench ranks them by the job. Here are the five seats in our current rotation, each earned by one kind of work, each held only as long as the work gets done.

TLDR

Five seats, five jobs, all sourced from the vendors' own July posts. A seat is a working assignment, not an endorsement.

•  Gemini 3.6 Flash holds the fast lane: drafts, routing, and repeated processing.

•  Gemma 4 12B holds the local seat: private work that never leaves the laptop.

•  Stable Audio 3.0 holds the sound seat: open-weight audio sketching.

•  Ideogram 4.0 holds the lettering seat: words that must survive inside the image.

•  Stable Diffusion 3.5 with TensorRT holds the volume seat: batch renders on local hardware.

The question we get most often is some version of which model is best, and the honest answer is that the question does not survive contact with real work. A workflow does not need the best model. It needs the right model for each seat, the way a small crew needs a gaffer more than a second director.

So this issue is our bench as it stands this week: five seats, five jobs, and what each model has to keep doing to hold its place. Every claim below traces to the vendor's own July announcement, and we treat vendor numbers as vendor numbers.

Google DeepMind introduced Gemini 3.6 Flash alongside 3.5 Flash-Lite and 3.5 Flash Cyber on July 21. In our routing, Flash-class models take the repeated passes: first drafts, summaries, classification, the second set of eyes that runs fifty times a day. The job is throughput with acceptable judgment, not brilliance.

The seat rule from issue 127 still applies: difficult judgment and repeated processing get different lanes. Flash keeps this seat as long as its mistakes stay cheap and catchable at the gate.

Google's developer blog walked Gemma 4 12B onto an ordinary laptop through AI Edge, aimed at local agentic workflows. The seat exists for one reason: some client material should never leave the machine. Contracts, unreleased work, anything under an NDA that a cloud terms-of-service page cannot soothe.

The local seat trades ceiling for custody. Drafts come out rougher and the second pass matters more, but the file never traveled, and for a working studio that sentence closes the debate.

Stability describes Stable Audio 3.0 as a model family built for artistic experimentation with open weights. That framing matches the job: sound sketching, not sound delivery. Mood beds under a rough cut, texture tests, the audio equivalent of a thumbnail pass.

Open weights earn the seat because experiments multiply. When a direction dies, nothing is owed; when one lives, the final track is still a musician's decision, made with better references than silence.

Issue 126 covered why in-image text is its own discipline, and nothing since has unseated it: when the words must survive inside the picture, posters, labels, title cards, the lettering model is chosen by whether the type holds, not by how the clouds look.

Ideogram's open-model inference is now scaling on AMD hardware per AMD's July post, which matters to us only as a signal of staying power. The seat is held by one test: spell the headline, ten times out of ten.

Stability and NVIDIA report Stable Diffusion 3.5 optimized with TensorRT at roughly twice the speed and forty percent less memory on RTX cards. Vendor numbers, so hold them loosely, but the seat itself is real: somebody has to render the eighty variations nobody will ever see.

Volume work belongs on hardware you already own, where a failed batch costs electricity instead of credits. The keeper images then graduate to whatever finishing pass the client brief demands.

Every source above is a vendor writing about its own model, and every seat was earned inside our particular work. Your bench should disagree with ours somewhere, or it is not yours. Review the seats monthly, evict without sentiment, and let no model keep a chair on reputation.

Best-model debates are entertainment. Name the five jobs your work actually contains, then audition models for one seat at a time. The bench you end up with will be smaller, cheaper, and calmer than the discourse, which is the point.

THE NEXT WORKING NOTE

I am writing up the bench card: a one-page worksheet for naming your five seats, the audition test for each, and the monthly review ritual.

Want it when it ships? Reply with send me the bench card and I will get it to you.

A QUESTION FOR YOU

Which seat on your bench is empty right now?

Reply with the job in your workflow that still has no reliable model assigned to it. The most common empty seat becomes a future audition issue.

If this was useful, forward it to a creator who is turning prompts into a repeatable system.