A generated title card headed 'Karpo Decision 31B vs Drex Max 1.1' with the subheading 'Two boards, one number each — and no head-to-head', two rounded panels side by side labelled 'Decision Index' reading 'Karpo Decision 31B 73.9' and 'Artificial Analysis v2.1' reading 'Drex Max 1.1 72.62', and a small vertical chip in the gap between them reading 'no shared instrument'.
Guides & Insights

Karpo Decision 31B vs Drex Max 1.1: Two Boards, One Number Each

Author

Elias Hawthorne

Date Published

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

Karpo Decision 31B and Drex Max 1.1 are close enough in size, shape and job description that a spec sheet makes them look like the same product from two companies. Both are dense ~30B models fine-tuned to answer typed questions about a state in one forward pass without generating text. Both ship a 255-way readout head, both accept images, both refuse rather than truncate an oversized question. The divergence is not in the architecture. It is in what each vendor has been willing to let someone else measure, and in what you are allowed to do with the weights afterwards. Karpo, from MachinePulse, has a third-party Full score and no price. Drex Max 1.1, from Nace.AI, has an independent Artificial Analysis number, a published rate and a licence that restricts commercial use.

Neither of those trades is obviously the right one. This is what each side of the trade actually costs.

Where the numbers come from, because they do not come from the same place

Start with the uncomfortable part: on no single leaderboard do both models have a score today.

Karpo Decision 31B appears on the Decision Index, the community board tracking open reproductions of TypeSafe's Je​v. Its Full score is 73.9, computed by the board's maintainers on private tests, and it currently shares first place with GLiDE 27B at 73.4. Its public index of 75.05 came from MachinePulse's own run of the official kit and was adopted by the board, with the 140,178-request results file published for anyone to re-run. Artificial Analysis has no row for Karpo at all.

Drex Max 1.1 is the mirror image. Artificial Analysis scores it at 72.62 on Decision Index v2.1 — second of 25 decision models, behind Liquid AI d1.2-ultra at 74.85 and ahead of SubQ Snap 1.2 at 72.58. That number is AA's, measured on AA's hardware with AA's private document set. What Karpo has, Drex Max 1.1 does not: the Je​v tracker carries a row for Drex Max v1 at 66.60 Full, and no row at all for v1.1.

So the only same-instrument comparison available is Karpo Decision 31B at 73.9 against Drex Max v1 at 66.6 — a 7.3-point gap on one board, across two different versions of the Nace model, which is exactly the kind of comparison that flatters whichever side you happen to favour. Draw the line the other way and Drex Max 1.1 has a 72.62 where Karpo has nothing printed. Both sentences are true. Neither is a head-to-head.

The specs, dimension by dimension

A generated two-column scoreboard titled 'Karpo Decision 31B vs Drex Max 1.1 — the scoreboard'. The Karpo column reads 'Params: 31.3B, 62.5 GB bf16', 'Context: 32,768 tokens', 'Score: 73.9, Decision Index', 'Price: none published', 'Licence: Apache-2.0' and 'Hosting: self-host only'; the Drex Max 1.1 column reads 'Params: dense 27B, 49 GB bf16', 'Context: 16,384 default, up to 131,072', 'Score: 72.62, Artificial Analysis', 'Price: $0.03 per million input', 'Licence: Nace.AI Open RAIL-M' and 'Hosting: served by Nace.AI'. The footer reads 'Karpo figures per the Decision Index and its model card; Drex figures per Artificial Analysis and Nace.AI.'

• Parameters — Karpo 31.3B in bf16 plus a 255 × 5,376 fp32 readout, about 62.5 GB on disk; Drex Max 1.1 is a dense 27B, about 49 GB in bf16 and 26 GB with the backbone at 8-bit.

• Base model — Gemma 4 31B instruction-tuned for Karpo, Qwen3.8-27B for Drex Max 1.1. Different pretraining corpora, different tokenizers, different everything downstream.

• Context — Karpo takes 32,768 tokens per sequence and splits a longer request by question, each part carrying the full state; Drex Max 1.1 defaults to 16,384 tokens per request and the vendor's documentation raises it to 131,072 per question.

• Multimodality — both read images natively, and both make the screenshot visible to every question in the request rather than to one.

• Question types — identical surface: choice up to 255 options, noul for true/false, score for an ordered level.

• Interface — Karpo serves System One at POST /v1/systemone from its own Python package; Drex Max 1.1 has the same shape, plus llama.cpp and Ollama forks for Apple silicon and CPU.

• Licence — Apache-2.0 for Karpo's weights and code; Nace.AI Open RAIL-M for Drex Max 1.1, which is open weights with use restrictions attached.

• Hosting — Karpo has none. Drex Max 1.1 is served by Nace.AI at $0.03 per million input tokens, flagged on the AA board as a temporary launch price.

A capture of the community Decision Index leaderboard, headed with the multimodalart jev-decision-index attribution, listing Karpo Decision 31B at 73.86 in first place ahead of GLiDE 27B at 73.37, REAX 1.3 at 72.64, Logits Sigma-1 Duo at 72.28, Copernicus-31B at 72.11, KnowLine 31B Gen1.5 at 71.91, Copernicus-57B at 71.8, FlexDM-LoRA 2.1 at 71.66, Sieve 31B v3 at 71.10, OneJev1.6g-31B at 70.86 and Turion Jev 31B at 70.47.

The context row is the one people skim past. A 32,768-token hard ceiling with an explicit refusal is a different engineering constraint from a 16,384-token default you can raise to 131,072 with an environment variable. If your states are invoices and support tickets, neither number matters. If your states are contracts, the Nace model can be handed ten of them and Karpo cannot be handed one.

What 11 points of calibration difference looks like in a pipeline

Accuracy is the number people quote and the number that matters least for this class of model. The whole point of a decision model is that its output is a probability a threshold can act on, so the question is whether the probability is honest.

Karpo's card reports a calibration error of 4.8 points: when the model says 0.9, it is right somewhat less than 90% of the time, by about five points on average. Drex Max 1.1's AA profile carries a Brier score of 0.16 across the full question set, which is a different statistic and not directly comparable, but the same board also shows its confidence ladder: at the 0.98 confidence bucket the model is right 97.5% of the time. That is a model whose top-confidence answers can be trusted at face value.

The interesting split for Drex Max 1.1 is in the knowledge evaluation. On AA-Omniscience it scores a decision score of 60.3 — but that composite comes from 28% raw accuracy against a non-hallucination rate of 65.3%. Read those together and what you have is a model that answers about a quarter of knowledge questions correctly and declines to answer a large share of the rest. The index pays it for declining rather than inventing, which is the intended behaviour for a decision layer and a poor fit for anything that needs an actual answer.

For Karpo the same instinct shows up in a different place: nine requests out of 140,178 were refused outright because a question plus its state overflowed the window. Refusal over silent truncation is the right default. It is also a failure mode you have to route around, and a 32,768-token ceiling makes it a more frequent one.

Latency and the shape of the deployment

Karpo's recorded median is 94 ms per request on the tracker's reference hardware, one GPU, one request at a time. AA measures Drex Max 1.1 at 283 ms per task and a cost of $102.7 per million tasks at the current launch price. The two numbers are not the same measurement — per request against per task, and a task on AA's document set is several decisions over one long document — so treat the gap as directional rather than precise.

Directionally it says something real: a 31B Gem​ma-derived readout with a single shared state pass is a leaner object than a 27B model whose non-causal attention layers rewrite the whole sequence. Whether leaner matters depends on where the call sits. A gate that fires before every tool call needs to be inside the turn budget. A nightly re-scoring of a thousand tickets does not care about 200 ms.

The licence question nobody asks until procurement does

Karpo Decision 31B is Apache-2.0, weights and code, with Gem​ma 4's own licence carried alongside in the repository. That is the permissive end of open.

Drex Max 1.1 is under the Nace.AI Open RAIL-M licence. Open RAIL-M is genuinely open — you can download, run, modify and redistribute — but it carries use restrictions that flow to downstream users, and it obliges you to pass the same restrictions on. For a research notebook that distinction is invisible. For a product where the model sits inside a customer-facing pipeline, it is a legal review, and the review will ask what happens when you fine-tune it and ship the derivative. The answer is "the restrictions come with it".

If the decision layer is internal tooling, the licence is a footnote. If it is embedded in something you sell, Apache-2.0 removes a meeting from the critical path.

Which one actually fits

Take Drex Max 1.1 if you want a number you did not have to compute, a price you can put in a budget, and states that run long. It is measurable today, it is hosted today, and its confidence ladder at the top end is the property that makes an automated threshold defensible. The costs are the licence restrictions and a launch price that AA explicitly flags as temporary.

Take Karpo Decision 31B if the absence of a hosted endpoint is not a problem — you have the GPU, or you have a batch job — and the permissiveness of Apache-2.0 matters, and the board's Full score is the evidence you trust. The costs are 62.5 GB of weights, a 32,768-token ceiling, and a vendor that has not published a price, a product page or an announcement.

There is a third position, and for most teams it is the correct one this month: neither. Karpo is two weeks old with a tied lead and no independent score; Drex Max 1.1 is priced on a discount that expires. Run both against your own labelled sample and compare the confident-and-wrong rate rather than the headline. For the incumbent in that shadow test, the cheapest available reference is TypeSafe's own model — Jev 1.14 is on our catalogue, so a threshold policy can be validated against the thing every open reproduction is chasing, on the same key you would use for Qwen3.8-27B or Gemma 4 31B, the two base models these checkpoints were built from. Both bases are served through one API for 200+ models at provider list price with no markup, which is useful precisely because a decision layer is a place you want to swap the model and keep the code.

The single thing that would settle this matchup is one number: Karpo Decision 31B appearing on the AA board. Until then the comparison is a well-documented 73.9 against a well-documented 72.62, produced by two organisations that have never graded the same paper.

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.