A generated title card headed 'Karpo Decision 31B' with the subheading 'A 31B Gemma 4 fine-tune at the top of a 307-model board, with no launch post to its name', three chips reading 'weights: 5 November 2026', 'Full score: 73.9' and 'announcement: none', a four-point horizontal timeline reading '5 Nov repository', '6 Nov scoring run', '12 Nov six-model family' and '16 Nov README edits', and an outlined card reading 'no price', 'no endpoint' and 'no Artificial Analysis row'.
Guides & Insights

Karpo Decision 31B Tops a 307-Model Board With No Launch Post to Its Name

Author

jinhao.song

Date Published

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

On 12 November 2026 the Decision Index — the community leaderboard that tracks open reproductions of TypeSafe's Je​v — put Karpo Decision 31B at the top of its Full score, 73.9, the highest of 307 entries. That same week the claim began circulating that this model, a fine-tune of Goog​le's open Gemma 4 31B, beats Goog​le's own Gem​ini Decision 1. Only one of those two statements is something a reader can check today, and the difference between them is the whole story. The board placing Karpo first is public, dated and reproducible from a 45 MB results file the lab published. The head-to-head against Gem​ini Decision 1 has never been run: the two models do not share a single leaderboard, and the instruments they were scored on are not the same instrument.

What makes the first half strange is not the score. It is the silence around it. MachinePulse Pte. Ltd., the Singapore company behind Karpo, created the Hugging Face repository on 5 November 2026, published an inference repository on GitHub the same day, and shipped five smaller siblings — 12B, 9B, 4B, 2B and 0.8B — on 12 November. As of this writing there is no product page, no blog post, no pricing, no model announcement and no API. The company's own site returns a 404 on every path that would hold one. Karpo exists as weights, a README, a Python package and a leaderboard row. This piece is about what those four things can and cannot tell you.

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 dates, and why they matter more than usual

The repository creation stamp is 5 November 2026 at 15:33 UTC. The GitHub inference code was created twenty-two minutes later. The Decision Index run dataset — the thing that makes the 73.9 auditable — was uploaded on 6 November at 05:18 UTC. The README was still being edited on 16 November, the last two commits changing nothing but a copyright line and a licence statement.

That is a two-week-old model with an active repository and no launch. It is not a leak, and it is not a rename of something older: the weights are new, the licence file is new, and the six-model family was pushed in a single afternoon. What it is, is a lab that decided to publish the artifact before the marketing.

What the 73.9 measures, precisely

A single-column scoreboard titled 'Karpo Decision 31B — the scoreboard' with six rows reading 'Full score: 73.9, third-party private tests', 'Public index: 75.05, vendor run, board-adopted', 'Rank: joint first of 307', 'Artificial Analysis: no row', 'Price: none published' and 'Licence: Apache-2.0, Gemma 4 terms alongside', footed 'Full score computed by the Decision Index maintainers; public index from a 140,178-request vendor run adopted by the board.'

The Decision Index as it stands is edition 0.4.1. Its Full score is computed by the board's maintainers on private tests: roughly 60% tasks drawn from the same skills as the public suite and 40% from domains the public suite does not cover. Karpo's row reads 73.9 Full, with 70.7 on the same-skills portion and 78.6 on the new-domain portion. Those last two numbers run in an unusual order — a model usually does worse on unfamiliar territory than on the skills it was tuned against — and the card reports it without comment.

The board publishes one rank to models within 0.9 points of the leader, so Karpo does not hold first place alone. GLiDE 27B, at 73.4, shares the #1 slot. Behind them sit REAX 1.3 at 72.6, Logits Sigma-1 Duo at 72.3, Copernicus-31B at 72.1 and KnowLine 31B Gen1.5 at 71.9. This is a crowded top, not a runaway.

The second number on the card, 75.05, has a different provenance and deserves to be read differently. That is the public index, and it comes from MachinePulse's own run of the official scoring kit — 140,178 scored requests, nine of them refused because a single question plus its state exceeded the 32,768-token window. The board adopted that run. The reason to trust it more than a typical vendor figure is that the run directory is public: results.jsonl.gz, a per-request scores file, an environment manifest. Anyone can re-run the kit and compare. The board's own adoption is the review step. That is a stronger standard than most launch claims meet, and it is still not the same as an independent lab scoring the model on its own hardware.

Two labels, then, and they should not be collapsed: the Full score is third-party-computed from private tests; the public index is first-party-computed from a public suite and adopted. Neither is a vendor benchmark in the usual sense. Neither is Artificial Analysis.

Why "beats Gem​ini Decision 1" is not yet a measured statement

Gem​ini Decision 1 has a number. Artificial Analysis scored it at 65.98 on Decision Index v2.1, a composite of six evaluations weighted across document decisions, tool calling, knowledge, reasoning and invariance. On that board it sits eleventh of the decision models AA ranks, below d1.2-ultra at 74.85, Drex Max v1.1 at 72.62 and SubQ Snap 1.2 at 72.58, and above Microsoft-Decision-1 at 62.07. The scoring is AA's, on AA's hardware, and it costs $345 per million tasks to run.

Karpo does not appear on that board at all. Artificial Analysis has no page for it and no row for it. Nor does Gem​ini Decision 1 appear on the Decision Index tracker that carries Karpo — that board is a register of open reproductions, and Goog​le's model is a closed, API-only product with no weights to reproduce.

So the comparison everyone is making is an inference from two numbers produced by two organisations on two suites with different task mixes, different private-test shares and different scoring formulas. 73.9 is not bigger than 65.98 in any sense that survives contact with the methodology. If the two were run through the same kit tomorrow, Karpo might well win — its per-category profile is strong — but nobody has run it, and the headline is currently a hypothesis wearing a result's clothes.

There is one further wrinkle worth knowing, because it cuts against the claim rather than for it. On the tracker's five category breakdowns, Karpo's weakest area is knowledge and reasoning, at 65.5 chance-corrected skill. Its strongest is tools and automation, at 88.4. On AA's board, the single component where Gem​ini Decision 1 beats almost everything is the knowledge one: 66.3 on AA-Omniscience, the highest of the four models checked here. The area where the Goog​le model is best is the area where the Gem​ma fine-tune is comparatively weakest. A head-to-head would be a genuine contest there, not a walkover, and that is precisely why the unrun comparison should not be reported as a win.

The part of the card that is more interesting than the ranking

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.

Karpo is not a chat model with a JSON mode. It is a readout. Each option in a question gets an option code, and a linear head scores 255 codes at a fixed slot, then softmaxes over the options the question actually offered. Nothing is generated, so the answer is always one of the choices you supplied and the output cost is structurally zero.

The architecture is worth one paragraph because it explains the latency. A request is one state plus any number of questions, encoded as a single sequence: questions first, then the state, then one branch per question. Branch attention lets a branch read the shared block and itself, never another branch. The state is computed once. A hundred questions about one invoice therefore cost roughly one pass over the invoice, not a hundred.

• Parameters — 31.3B in bf16, plus a 255 × 5,376 readout in fp32; about 62.5 GB on disk.

• Context — 32,768 tokens per sequence; a longer request is split by question, each part carrying the full state, and a question that still will not fit is refused rather than truncated.

• Input — text, JSON and images; a screenshot can be passed as state.viewport_screenshot or as a data URL.

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

• Interface — a Python API and an HTTP server in the System One shape, POST /v1/systemone.

• Licence — Apache-2.0 for the weights and the code, with Gem​ma 4's own licence carried alongside in the repository.

The measured latency on the tracker's reference hardware, one GPU and one request at a time, is 94 ms median. That is not fast enough to be invisible inside a per-step agent loop that makes twenty small calls per turn, and it is fast enough to sit behind a tool-call gate that fires a handful of times per conversation. The card also reports a calibration error of 4.8 points, which is the number a deployment actually hinges on: accuracy tells you how often the top answer is right, calibration tells you whether the probability attached to it can be thresholded.

What is not confirmed, stated plainly

There is no price. Not "undisclosed pending launch" — no rate card, no hosted endpoint, no console, no trial. If you want Karpo Decision 31B today you download 62.5 GB and run it on your own GPU with the repository's own loader. Loading the checkpoint as an ordinary chat model will not reproduce its answers; the branch attention and the readout are both required.

There is no independent score. The Decision Index is unofficial and community-maintained. It is careful, well-documented and unusually transparent about method, and it is not Artificial Analysis, and it has not been asked to score this model on its own hardware. The strongest third-party element is that its maintainers computed the Full score on tests the vendor never saw.

There is no vendor statement of what the model was built for beyond the README's own description: the small calls an agent makes at every step — whether to call a tool or ask first, which element on a page to click, whether a return is eligible under a policy. That description is the lab's framing, not an independently validated scope.

And there is no announcement. Not a quiet one, not a changelog entry. Two weeks after the weights went up, the company's public surface is a hiring page and a different product line.

Where this leaves a team that wants to try it

The honest summary is that Karpo Decision 31B is the most interesting open decision model of the past fortnight and the least supported one. Its lead is real but tied, its public index is reproducible, its private-test score is genuinely third-party, and its central marketing claim — that a Gem​ma fine-tune now beats Goog​le at its own decision model — is unmeasured.

For teams already running a decision layer, the practical route is a shadow deployment rather than a migration. Keep the model that is carrying production traffic, run Karpo on the same requests offline, and compare the probabilities rather than the labels: a 4.8-point calibration error means roughly one answer in twenty is more confidently wrong than its stated probability suggests, and your threshold policy is what turns that into either a human review queue or an incident. If you are standardising on one endpoint for the rest of the stack, the models you can actually reach this week are the ones on the catalogue — Gemma 4 31B itself, which is Karpo's base and is served through one API for 200+ models at provider list price with no markup, and TypeSafe's Jev 1.14, the model this whole open-reproduction board was built to chase. Karpo will have to earn its slot the way the others did: by being callable.

The thing to watch is not the 73.9. It is whether MachinePulse ships a product page, a price and an AA submission in the next month. A lab that publishes weights before a website usually has a reason, and the reason tends to show up as either a hosted endpoint or a pivot. Either way, the first independent number on the AA board will settle in a week what a fortnight of leaderboard arithmetic has not.