Ollama's new endpoint puts typed routing and moderation decisions on the local machine

A call to /v1/systemone carries a `state`, which can be text or structured data, and a set of named questions ^[2]. Each question declares the kind of answer it wants: a choice, a yes-or-no, or a score ^[2]. Ollama's example sends one support ticket with three questions: which team should own it, whether it requests a refund, and how urgent it is ^[4]. The reply puts each answer in a defined field, with a probability for every allowed choice ^[3]. Routing code reads the field and branches. It does not parse generated text ^[3].

This is good engineering for narrow jobs. A dispatcher needs a closed answer set and a number it can compare, and the endpoint returns both ^[3]. With Nimble's limit of 64 questions per input ^[6], one call can carry a full triage checklist. The endpoint also follows the API shape of Jev, TypeSafe's hosted decision model ^[11].

The 91 ms average comes from a Pac-Man demonstration ^[8], and a ghost turning the wrong way costs less than a wrongly blocked customer. Ollama measured it on one MacBook Pro with an M5 Max ^[8]. For the figure to transfer, the target machine needs comparable memory and compute. The local route removes network transit and per-request hosted charges, and the hardware has to cover what is left ^[16].

The accuracy figures are Bespoke Labs' own, reproduced on Ollama's Nimble page ^[10]. They cover 3,880 examples across 13 public datasets ^[10]. That is about 298 examples per dataset on average ^[4]. Across all of them, Nimble trails Jev by 1.2 points of macro-average accuracy ^[2]. On Civil Comments, the moderation set, Jev agreed with human labels 81.0% of the time ^[12]. There the gap widens to 10.7 points ^[1]. Nimble disagreed with the annotators on 29.7% of those examples ^[3].

According to Runtimewire, these are agreement scores against dataset labels and do not prove performance on a company's own moderation queue ^[17]. For the Civil Comments number to hold, a company's content would have to resemble that dataset and its policy would have to match those annotators. I would not assume either for a support desk or a marketplace.

Nimble also returns a confidence measure. Ollama's documentation says it reflects how concentrated the model's choice is and does not promise the answer is correct ^[13]. A sharp distribution can still point at the wrong label. In my view the human-review cutoff has to come from labelled items in the real queue. I'd run a few hundred through Nimble, sort the errors by confidence, and draw the review line where the error rate stops being tolerable. Because the endpoint follows Jev's API shape ^[11], the same sample should run against the hosted model with little rework.

The launch post presents local use as carrying no additional inference cost, and Ollama has not published a separate price for the endpoint ^[14]. Ollama raised a $65 million Series B led by Theory Ventures in July, bringing its total funding to $88 million ^[15].