A generated title card headed 'Karpo Decision 31B vs Liquid AI D1 Ultra' with the subheading 'One has a price. The other has a story.', four chips reading '$0.042 per million input tokens', '70.58 on Artificial Analysis v2.1', '73.9 on the community board' and 'no Artificial Analysis row for Karpo', and two outlined cards separated by a thin divider labelled 'priced and served' and 'open weights, no endpoint'.
Guides & Insights

Karpo Decision 31B vs Liquid AI D1 Ultra: One Has a Price, the Other Has a Story

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Liquid AI D1 Ultra is the only model in this comparison that a finance team can approve. It shipped on 15 October 2026, it is served from Liquid's own infrastructure at $0.042 per million input tokens with zero output tokens by construction, and Artificial Analysis scores it at 70.58 on Decision Index v2.1 — eighth of 25 decision models, 4.3 points behind its own successor d1.2-ultra and roughly level with Drex Max 1.1 at 72.62. Karpo Decision 31B, from MachinePulse, has no price, no endpoint and no Artificial Analysis row. What it has is a 73.9 Full score on the community Decision Index, computed by that board's maintainers, and a two-week-old Apache-2.0 repository that nobody has announced.

Set side by side, the two models describe the two ways a decision model can enter an organisation: one arrives with an invoice and a measured number, the other arrives with a download link and a promise. Which you should care about depends on whether the decision layer is going into production this quarter or being prototyped against next quarter's budget.

The number Liquid is actually selling, and why it is the non-hallucination one

A generated two-column scoreboard titled 'Karpo Decision 31B vs Liquid AI D1 Ultra — the scoreboard'. The Karpo column reads 'Score: 73.9 Decision Index', 'Price: none published', 'Cost per million tasks: self-hosted only', 'Latency: 94 ms per request', 'Non-hallucination: not reported' and 'Licence: Apache-2.0'; the D1 Ultra column reads 'Score: 70.58 Artificial Analysis', 'Price: $0.042 per million input', 'Cost per million tasks: $133', 'Latency: 301 ms per task', 'Non-hallucination: 88.6' and 'Licence: not open, API only'. The footer reads 'Karpo figures per the Decision Index; D1 Ultra figures per Artificial Analysis v2.1.'A capture of the Artificial Analysis Decision Index v2.1 page showing the six weighted evaluations — AA-Paperclip at 30%, EvalSafe-O*NET-AA at 15%, When2Call at 10%, AA-Omniscience-Decision at 15%, HLE-MultiChoice at 15% and AA-Symmetry at 15% — the methodology note that all evaluations are conducted independently by Artificial Analysis, the 'Update Oct 30, 2026' stamp, and the worked example of a support-ticket state scored by a decision model.

D1 Ultra's headline is a first-place claim from October that the published board did not support, and that argument has been made elsewhere. The thing worth understanding now is what changed between it and its predecessor, because the AA data shows it cleanly.

Liquid AI d1, the original, scores 3.3 on AA's non-hallucination rate for the knowledge evaluation. That is close to the bottom of the entire board — a model that, when it does not know, invents. D1 Ultra scores 88.6 on the same measure. On the multiple-choice knowledge exam, it moves from 5.6 to 56.9. Almost none of that movement comes from getting more answers right: raw accuracy on AA-Omniscience actually fell from 39.8% to 28.0%. The model got better by learning to decline.

That is not a trick. For a decision layer it is the single most valuable behaviour there is, because a declined answer routes to a human and a confidently invented one routes to a customer. But it reframes what the index is rewarding. D1 Ultra's 70.58 is substantially a score for abstention, and the same board shows the cost: on HLE-MultiChoice, raw accuracy is 6.09%, one of the lowest figures of the 25 models listed.

The successor, d1.2-ultra, went further in the same direction and further past it — 83.8 non-hallucination with only 23.6% raw accuracy, an index of 74.85, and the top of the board. Liquid has been optimising one axis hard, and the axis is honesty rather than knowledge.

Where Karpo's profile runs the other way

Karpo's numbers come from a different instrument, so they cannot be subtracted from Liquid's. What they can do is show a different shape.

On the Decision Index's five categories, Karpo's strongest area is tools and automation at 88.4 chance-corrected skill, and its weakest is knowledge and reasoning at 65.5. Its tracker-recorded calibration error is 4.8 points. The card also reports a median latency of 94 ms on reference hardware, against 301 ms per task for D1 Ultra on AA's measurement — different measurements, one per request and one per task, so directional only.

Read the two profiles together and a coherent difference emerges: Karpo is a model tuned toward action — tool selection, routing, page interaction, the small calls an agent makes at every step — while D1 Ultra and its successor are tuned toward refusal. A model that declines 72% of knowledge questions is a poor tool-choice engine, because tool selection is not a question you are allowed to abstain from. A model that answers 88% correctly on tool routing but only 65% on knowledge is a poor knowledge engine.

That is the real matchup, and it is a matchup of intended use rather than of scores. MachinePulse's own framing in the model card says as much: the model is for the decisions an agent makes between steps, whether to call a tool or ask first, which element to click. Liquid's October release, and even more its November successor, is for judging whether a claim is supported.

The cost comparison, with the caveats attached

• Liquid AI D1 Ultra — $0.042 per million input tokens, no output charge because nothing is generated, $133 per million tasks on AA's document set, median 301 ms per task.

• Karpo Decision 31B — no published price. Self-hosted only: 62.5 GB of bf16 weights, one GPU, 94 ms median per request on the tracker's reference hardware.

• Drex Max 1.1, for scale — $0.03 per million input tokens at a launch price AA flags as temporary, $102.7 per million tasks.

• The comparison nobody publishes — a 62.5 GB resident model on a rented GPU costs whatever that GPU costs per hour whether you send it one request or ten thousand. At small volumes that is far more than $133 per million tasks; at high sustained volumes it inverts.

The per-task figure is the one to budget against, because it folds in how many tokens each model burns to answer the same document, and a model that repeatedly re-reads its state costs more per decision than its per-token rate suggests. D1 Ultra's $133 is the highest of the three Liquid models listed here and $30 above Drex Max 1.1's launch price. It is not a cheap decision layer, and the model it replaced was $63.4.

Where a hosted tier genuinely wins is at the low end. If you need a decision model for a hundred requests a day, provisioning 62.5 GB of memory for Karpo is not a serious option, and calling a served endpoint at list price is. Our catalogue passes provider list price through with no markup, so a Liquid price cut lands the same day it is announced, and a decision model can sit alongside the rest of the stack on one key rather than as a second contract. That is not an argument for D1 Ultra over Karpo; it is an argument for not buying a GPU to find out whether you need one.

What D1 Ultra can tell you that Karpo cannot

Three things, all of them the kind of thing you discover in month two.

The first is the confidence ladder. AA publishes per-bucket calibration for every model it scores, and D1 Ultra's is legible: in the 0.96-plus confidence bucket it is right 97.1% of the time on the symmetry evaluation. That tells you a threshold at 0.95 behaves roughly as advertised, which is a claim you can build a review queue on. Karpo's card gives one aggregate figure — a 4.8-point calibration error — and no per-bucket curve. You can produce one yourself, and you will have to.

The second is invariance. AA-Symmetry reorders options, swaps labels and negates yes/no questions across 440 variants, and D1 Ultra scores 78.7 with a 45.8% all-pass rate. A decision model that gives a different answer when you alphabetise the options is not a decision model, it is a coin with extra steps. Karpo's card does not report an invariance result, and because its readout scores option codes at fixed slots rather than option text, this is exactly the property worth verifying before deployment.

The third is behaviour on refusal. AA reports 88.6 non-hallucination on the knowledge set for D1 Ultra, which means the abstention path is exercised and measured. Karpo refuses only on length — nine of 140,178 requests, all oversized — and its card says nothing about abstention as a decision. Whether Karpo declines a question it cannot answer, or picks the least-bad option because the readout always picks something, is not documented. The limitations section says only that it "always picks one of the given options".

That last sentence is the most important one in either model card. A decision model that structurally cannot abstain needs a "no confident answer" option in every choice list, or a threshold on the probability, or both. D1 Ultra's designers chose the other architecture: it can decline, and it declines often.

The licence, which is the one place Karpo wins outright

Karpo Decision 31B is Apache-2.0, weights and code, with Gemma 4's licence carried alongside. Download it, run it, modify it, ship the derivative, no conversation with a lawyer required.

Liquid's models are not open. There are no weights, no self-hosting, no fine-tuning on your own traffic, and no air-gapped deployment. The API is the product. For an internal classification pipeline handling regulated documents, that is a procurement blocker in a way that a 62.5 GB checkpoint is not.

So the trade is stark and it is not about scores: Karpo gives you control and nothing else; Liquid gives you a measured product and takes away the weights. A team with a compliance requirement cannot choose D1 Ultra at any price. A team with a three-week deadline cannot choose Karpo at any effort, because there is no endpoint to call.

What to do this month

If the decision layer is going to production and the input is documents, D1 Ultra is the defensible choice despite the price and the low raw knowledge accuracy, because it is measured, served and priced, and its abstention behaviour is the property you actually want. Watch d1.2-ultra instead if you are buying now — it is the same product line, four points higher on the same board, and it carries the same caveat that a large share of its score is abstention rather than knowledge.

If the input is a page, a screenshot or a tool call and you have GPUs, Karpo is the more interesting model and the more permissive licence, and its tool-routing skill of 88.4 is the number to test against your own traffic. Go in knowing that you are the first person to measure it on anything that matters: no AA row, no price, no announcement, and a vendor whose product page currently returns four-oh-four.

There is a middle path worth naming, because it is what most teams end up doing. Run the hosted model in production, run the open one against the same requests offline, and compare the confident-and-wrong rate rather than the top-line index. That comparison needs both models callable on one credential and priced transparently, which is the case our catalogue exists to make — a served decision model and the open bases these fine-tunes were built from, on one key, at provider list price. If Karpo gets an endpoint or an AA submission in the next month, the latency of that change is now the thing to watch; if it does not, the 73.9 stays an interesting number on a community board and D1 Ultra stays the model with a receipt.

A capture of the Hugging Face model card for MachinePulseAI/karpo-decision-31B showing the Text Classification tag, the gemma4 image-text-to-text tag, the apache-2.0 licence, the line 'Ranked #1 on the Decision Index, with the highest Full score, 73.9', the description of a multimodal decision model that returns a probability for every option of every question in one forward pass without generating text, 'Fine-tuned from google/gemma-4-31B', 'Parameters: 31.3B in bf16, plus a 255 x 5,376 readout in fp32', the screenshot input at state.viewport_screenshot, and the base-model link to google/gemma-4-31B.