Discussion

Karpo Decision 31B and Liquid AI D1 Ultra take very different routes to AI decisions

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#14632

Karpo Decision 31B and Liquid AI’s D1 Ultra are aimed at different kinds of AI decision-making, according to a new comparison by OrcaRouter. The useful distinction is not a leaderboard winner: one is a self-hosted model with an open licence, while the other is a priced, hosted service with published benchmark results.

Watch Desk analysis

What happened

OrcaRouter compares the models for tasks such as tool selection and judging whether a claim is supported. It says Karpo has a 73.9 score on a community Decision Index, while D1 Ultra scores 70.58 on Artificial Analysis’s Decision Index v2.1. Those are different benchmarks, so the numbers are not a head-to-head result.

The comparison describes Karpo as available to download under Apache-2.0, with 62.5 GB of bf16 weights and no published price or hosted endpoint. D1 Ultra, it says, is available through Liquid AI’s service at $0.042 per million input tokens, with no output charge because the model returns decisions rather than generated text.

Key findings

  • Karpo is built for teams that can run it themselves
    OrcaRouter reports 94 ms median latency on reference hardware, but says the model needs a GPU and has no listed hosted price.
  • D1 Ultra has a measurable refusal profile
    The comparison says it scores 88.6 on Artificial Analysis’s non-hallucination measure for a knowledge evaluation, while raw accuracy on that test is much lower.
  • The scores measure different things
    Karpo’s community-board result and D1 Ultra’s Artificial Analysis figures cannot establish which model performs better on the same workload.

Why it matters

For a team choosing a decision layer, the practical question is whether it needs a model it can host and adapt, or one it can call as a service with published pricing and benchmark data. Karpo’s reported strength is tool routing; D1 Ultra’s is declining to answer when it lacks confidence. Those behaviours suit different jobs, and the comparison suggests testing them separately rather than treating “decision model” as one workload.

The evidence is not evenly matched. OrcaRouter says Karpo has no Artificial Analysis result and that its reported latency comes from different hardware and measurement conditions than D1 Ultra’s task-time figure. Treat the comparison as a useful guide to what to test, not a controlled contest.

Our read

The best takeaway is that “open versus hosted” is only half the decision. Teams should test whether a model can abstain appropriately, and whether its tool choices hold up on their own tasks. A score without a shared test is a conversation starter, not a procurement verdict.

What to watch

  • Whether Karpo gains a hosted endpoint or a comparable independent evaluation.
  • How both models perform on the same tool-routing and knowledge tests.
  • Whether Liquid AI changes D1 Ultra’s pricing or publishes further evaluation details.

Discussion spark: For an AI decision layer, would you prioritise control and self-hosting, or a hosted model with published pricing and benchmark results?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.