Most AI agents make dozens of small decisions per task: which button to click, which tool to call, whether to ask for clarification. Today they usually make them the slow way, by having a large language model write out an answer token by token. MachinePulse (registered as MachinePulse Pte. Ltd.) wants to change that with Karpo Decision, a family of six open-weight models that return a probability for every option in a single pass, without generating any text. Its largest model, Karpo Decision 31B, shares the top rank on the Decision Index leaderboard, and a wave of posts about it trended on X. Here is how the model works, what the ranking really says, and what to keep in mind about the hype around it.
Quick Take
- What: MachinePulse released Karpo Decision, six open-weight “decision models” from 0.8B to 31B parameters, under the Apache-2.0 license on Hugging Face.
- How: Instead of generating text, the model scores labeled options with a readout layer and returns calibrated probabilities, answering many questions about one page or document in one forward pass.
- Ranking: The 31B model scores 73.9 on the Decision Index 0.4.1 Full score, measured by the leaderboard maintainers on private tests. It shares rank #1 with GLiDE 27B (73.4).
- Speed: MachinePulse reports a 94 ms median latency per request on a single GPU.
- The catch: Results are self-published or come from one leaderboard, the model needs 62.5 GB of GPU memory, and much of the X buzz looks like coordinated promotion.
In plain English
AI assistants that browse websites or use apps have to make lots of tiny choices, like which link to click next. Normally they “think out loud” by writing text, which takes time. Karpo Decision is built to skip the writing and simply pick the most likely option from a list, with a confidence score, very quickly. Its makers have published it for free, and it ranks at the top of one specialist leaderboard for this kind of task.
What happened
MachinePulse published the 31B model and five smaller siblings on Hugging Face, along with inference code on GitHub and a product page on its Karpo website. Leah W, who builds the Karpo assistant at the company, announced the release on X:
Picking the next tool shouldn’t take a paragraph.
— LeahW (@LeahW_2077) November 16, 2026 · View on X
We built Karpo Decision. A family of six open-weight models, from 0.8B to 31B.Give it the context, a question, and the options. It returns a probability for each option in one forward pass.
The model card states that Karpo Decision 31B is “Fine-tuned from google/gemma-4-31B-it, with a linear readout over the options.” The 12B model is also based on Gemma 4, while the four smallest models (9B, 4B, 2B and 0.8B) are built on Qwen3.5. The weights are licensed under Apache-2.0, with Gemma 4 remaining under its own license.
Background: why decision models exist
Decision models are a small but growing category. We recently covered Microsoft-Decision-1.1, a small model built to route agent requests, and the broader case for decision models in agent architectures: many agent steps are multiple-choice questions, and paying a large generative model to write prose for each of them wastes time and money. The Decision Index, a public leaderboard with its own evaluation kit, ranks such models across tool use, retrieval, classification and judgment tasks. Karpo’s materials repeatedly compare against Google’s Gemini Decision 1.
How it works
According to the model card and Karpo’s product page:
- A readout, not generation: Every option gets a code. A linear readout scores up to 255 option codes, and a softmax over the question’s own options gives probabilities. “Nothing is decoded, so the answer is always one of the options.”
- Many questions, one pass: A request contains one state (a web page and screenshot, a document or a conversation) and any number of questions, encoded as one sequence.
- Branch attention: Each question’s branch can read the shared state and itself, but never another branch, so the state is computed once and questions do not influence each other.
- Three question types: “choice” (up to 255 options), “noul” (true or false) and “score” (a level on a described scale).
- Inputs and limits: Text and images are supported. The context is 32,768 tokens per sequence; longer requests are split by question, and a question that does not fit is refused, never truncated.
For example, a browser agent can ask, in a single request, which element to click, whether to call a tool and whether to ask the user for clarification about the same page.
What the numbers say
The Decision Index ranks by a Full score that the maintainers measure only on private tests: 60% tests of the same skills as the public suite and 40% tasks from new domains. On the board of November 12, the model card reports:
| Decision Index 0.4.1 (board of 12 November 2026) | Karpo Decision 31B |
|---|---|
| Full score (private tests) | 73.9, shared rank #1 of 300 entries |
| Private tests, same skills | 70.7 |
| Private tasks, new domains | 78.6 |
| Public index (own run, adopted by the board) | 75.0 |
| Median latency (one GPU, one request at a time) | 94 ms |
| Calibration error | 4.8 points |
The card notes that the board “gives one rank to models within 0.9 points of the first model in a group,” so Karpo Decision 31B and GLiDE 27B share #1. In her launch post, Leah W cites 87.2 on JevBench Capability versus 78.5 for Google’s Gemini Decision 1, and other posts about the release cite When2Call (90.4% versus 79.9%) and a New Yorker caption task (90.5 versus 79.7). One independent researcher on X highlighted that among the top public scorers, the model shows “the smallest drop on private same-skill tests (4.3 vs 6.8 points on average),” which suggests less overfitting to the public benchmark.
Pros and cons
For users
- Pro: Agents built on fast decision models can feel more responsive, especially when browsing or filling out forms.
- Con: End users will rarely see the model directly, and a confident wrong choice is still a wrong choice.
For developers
- Pro: Open weights under Apache-2.0, calibrated probabilities that support confidence thresholds, and a smaller family for cheaper hardware.
- Con: The 31B model needs 62.5 GB of GPU memory, uses a custom serving format, and “loading the checkpoint as a plain chat model does not reproduce its answers.”
For companies
- Pro: Self-hosting keeps sensitive page and document data in-house while cutting latency and cost for routine agent steps.
- Con: A young vendor and a single leaderboard are thin ground for production decisions; internal evaluation on real workflows is essential.
What X is saying
The release drew a large number of posts in a short time, many with near-identical wording (“Fast. Calibrated. Open.”, “tied #1,” “worth keeping an eye on”) across many accounts. That pattern does not say anything about the model’s quality, but it does mean raw post volume is not a useful signal here.
The more substantive reactions focused on methodology. Researcher Atal (@ZabihullahAtal) wrote:
From a research perspective, the Karpo-Decision-31B release is worth reading.
— Atal (@ZabihullahAtal) November 17, 2026 · View on X
It ranks #1 on the Decision Index (73.9 Full) and leads Gemini Decision 1 on When2Call, Mind2Web, New Yorker Caption and Amazon ESCI.
What I appreciate: the team reports overfitting openly. Among the top 10 public scorers, it shows the smallest drop on private same-skill tests (4.3 vs 6.8 points on average).
Some promotional posts also claimed a 256K-token context window, which contradicts the 32,768-token limit stated in the official model card. We rely on the model card. We did not find a Grok summary or reply about Karpo Decision at the time of writing.
Critical analysis
The core idea is sound and overdue: if an agent step is a multiple-choice question, a model that answers multiple-choice questions directly should be faster, cheaper and easier to calibrate than one that writes paragraphs. MachinePulse also deserves credit for a detailed model card, an Apache-2.0 license and openly reporting the gap between public and private test scores.
The caveats are about evidence and context. The headline “#1” is a shared first place under the leaderboard’s own grouping rule, on a single, specialized benchmark. The comparisons with Gemini Decision 1 are chosen by MachinePulse. The 94 ms latency was measured one request at a time on a high-end workstation GPU, which says little about throughput under load. And the promotional wave on X makes it harder to find independent hands-on reports. The model may well be excellent; the proof will come from developers plugging it into real agents and reporting what breaks, which is exactly what its makers have asked for.
What’s still open
- How the model performs in production agents beyond benchmark suites.
- Throughput and cost under concurrent load, not only single-request latency.
- Whether other decision models, including Google’s, publish comparable private-test results.
- How quickly hosting providers offer the model as an API (the Hugging Face page lists no inference provider yet).
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on the Karpo Decision 31B model card on Hugging Face, the Karpo Decision product page, and public posts on X by the developers and researchers, as of November 18, 2026. Benchmark figures are as reported by MachinePulse and the Decision Index board it cites.
Sources (plain text, no links):
- Hugging Face: MachinePulseAI/karpo-decision-31B model card
- Karpo: Karpo Decision product page (“Every answer in one forward pass”)
- GitHub: machinepulse-ai/karpo-decision (inference code, referenced in the model card)
- Posts on X by Leah W (@LeahW_2077) and Atal (@ZabihullahAtal) (November 16 and 17, 2026)
i10x Perspective
Agents do not need to be eloquent for every step; they need to be fast and right. Specialized decision models are a quiet but important trend, because they make agents cheaper and more predictable without waiting for bigger frontier models. Karpo Decision is one of the most complete open releases in this niche so far. Treat the leaderboard rank as an invitation to test, not a verdict, and watch the Agents section for how this category develops.






