Decision models, part two: train your own

This is part two. Part one, Decision models are popping up like mushrooms after the rain, is the problem. Fourteen decision models in four weeks, and every hosted one gets to see the decisions your agents make. This one is the solution.

What Coinbase did

Part one ended on the problem, that every hosted decision model sees what you were choosing between, and a few days later Brian Armstrong posted the solution I hadn't written yet. Coinbase fine-tuned an open-weight model, Qwen3.5-9B, on its own internal data for fraud prevention, and he says it "out performed frontier models, at much lower cost and speed" ^[1].

He thinks it's a pattern. "The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves." Coinbase made it easy enough that its staff start one from their internal agent harness with a single command, /fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review, and he expects some backend services to be "replaced entirely by this new architecture" ^[1].

He didn't publish the numbers, so I can't check them (and fraud numbers are the last thing a company publishes). But the pattern matches what's in our directory.

Same base model, different owner

Microsoft-Decision-1, which launched last Friday, is post-trained from the same Qwen3.5-9B ^[2]. Microsoft hosts it, you send it your questions and Microsoft sees them. Coinbase took the same starting point, trained it on its own cases and runs it in house, so the only one who sees which transactions looked like fraud is Coinbase.

That's part one's argument in a single base model. The model isn't the moat, the decisions are, and you get to choose who keeps them.

Most decision models are already open

The four hosted only are OpenAI's Decisions API, Jev, Celeris-1 and Microsoft-Decision-1 ^[3]. The other ten publish weights, most of them under Apache-2.0, and several are made to be trained further. Kev ships a kev-finetune skill that fine-tunes it on your own labels, and its README puts a Kev-4B training run at about a dollar on an H100 ^[4]. Laya has notebooks that train on two free Kaggle T4s ^[5]. Strands Decider is a small LoRA adapter and a head of about a million parameters on Qwen3.5-2B-Base ^[6], which is about as small as a decision model gets.

Clef and Kev also answer the same request as Jev's hosted API, so a team can start on a hosted model this week and move to its own fine-tune later by changing a base URL ^[7].

Why decisions are the easy thing to fine-tune

A decision is a narrow task with a fixed set of answers (fraud or not, approve or escalate, which of five queues). And the training data already exists, because every case a team has closed is a question with its answer next to it. A frontier lab doesn't have your closed fraud cases, and a 9B model doesn't need to know poetry to learn them.

It's also the work most companies already pay people and rule engines to do. A rules engine with years of if-statements in it is a decision model nobody trained, which I think is what Armstrong means about backend services.

The catch

A fine-tuned model is a thing you now own, which means you evaluate it, watch it and retrain it when the data moves. Laya's model card says its checkpoints come out over-confident and tells you to calibrate them on your own data ^[5], and that's the honest version of every fine-tune. A rule doesn't drift (it just gets wrong slowly, and someone notices eventually).

So I think most companies end up with both. A frontier model for the open-ended work, a hosted decision API to start with, and a few small models trained on their own decisions for the questions that come up thousands of times a day, once there's enough labelled data to train them.

Where to look

The decision model shortlist and every head-to-head have all fourteen, and each listing says whether its weights are open, under what licence and on which base model. If you train one on your own decisions, I'd like to hear how it went.

References

Read on 12 October 2026. Quotes are as published.

1. [1]Brian Armstrong on X, October 2026. Source for Coinbase's fine-tune of Qwen3.5-9B for fraud prevention, the quotes and the internal command. x.com
2. [2]Anchor Terminal, Microsoft-Decision-1. Source for its Qwen3.5-9B base and its launch, with Microsoft's pages. anchorterminal.com
3. [3]Anchor Terminal, decision models, 12 October 2026. Source for the fourteen models, which publish weights, their licences and their base models. Each listing links the pages it was graded from. anchorterminal.com
4. [4]Kev, README on GitHub. Source for the kev-finetune skill and the cost of a Kev-4B training run. github.com
5. [5]Anchor Terminal, Laya. Source for the Kaggle notebooks and the model card's advice to calibrate. anchorterminal.com
6. [6]Anchor Terminal, Strands Decider 2B. Source for the LoRA adapter, the head and the base. anchorterminal.com
7. [7]Anchor Terminal, Jev. Source for Clef and Kev answering Jev's request, so a client moves by changing its base URL. anchorterminal.com