Tuesday, November 17, 2026

Today’s Edition

AI Intel Report

MARKETS —

Frontier Models

Karpo Decision 31B tops Decision Index at 73.9 with 94ms latency

The open-weight model, fine-tuned from Google's Gemma 4 31B, returns calibrated probabilities in a single forward pass and ships with five smaller siblings on Hugging Face.

8 MIN READ
Close-up of a hand probing a circuit board in a lab with a magnifying lamp, representing testing of the Karpo Decision 31B model.
Illustration: AI Intel Report

Karpo Decision 31B is an open-weight multimodal decision model, fine-tuned from google/gemma-4-31B-it, that returns probabilities for typed questions — choice, noul and score — in a single forward pass without token generation.

Karpo released Karpo Decision 31B, an open-weight model that tied for No. 1 on the Decision Index 0.4.1 Full board with a score of 73.9, ahead of every other open-weight and proprietary model measured on the Nov. 12, 2026 ranking. The company also open-sourced five smaller models, from 0.8B to 12B, under an Apache-2.0 license, making the release one of the broadest open-weight decision-model families yet. The 31B flagship, fine-tuned from Google's google/gemma-4-31B-it, returns calibrated probabilities for typed questions in a single forward pass and reports a median latency of 94 ms per request on a single RTX PRO 6000 GPU (Hugging Face).

The release lands at a moment when decision models — systems that score options rather than generate text — are becoming a standard component in agentic stacks. Karpo's approach replaces token-by-token generation with a single forward pass, a design that changes the latency and cost profile for applications such as tool selection, retrieval ranking and workflow routing. The model card on Hugging Face describes the release as ranked No. 1 on the Decision Index, with a Full score of 73.9 and a median latency of 94 ms per request (Hugging Face).

What is Karpo Decision 31B?

Karpo Decision 31B is a multimodal decision model fine-tuned from google/gemma-4-31B-it. It is designed to answer typed questions — choice, noul and score — by returning a probability for each option in one forward pass, without generating tokens. The architecture allows the model to handle up to 255 options per question and a 32,000-token context, with longer contexts supported through splitting. Inputs can include images and screenshots, which makes the model applicable to UI-based agents that need to reason about what they see on screen.

The model is available on Hugging Face at MachinePulseAI/karpo-decision-31B under an Apache-2.0 license. Karpo said it open-sourced five additional sizes — 0.8B, 2B, 4B, 9B and 12B — so that developers can match model size to hardware and latency requirements. The full family is described on the Karpo Decision page, where the company says the 31B model posts a 73.86 Decision Index 0.4.1 Full score, ranking No. 1 (tied) (Karpo).

How does the Decision Index rank Karpo Decision 31B?

The Decision Index 0.4.1 Full board, published Nov. 12, 2026, ranks 300 models on private tests designed to measure decision quality. Karpo Decision 31B scored 73.9, tying with GLiDE 27B at 73.4 for the No. 1 slot. The Hugging Face model card notes that the board assigns one rank to models within 0.9 points, which is why karpo-decision-31B and GLiDE 27B share the top position (Hugging Face).

The score places Karpo Decision 31B ahead of Google's Gemini Decision 1, which did not top the Decision Index board but is a direct benchmark comparison on JevBench. On JevBench Capability, Karpo Decision 31B scores 87.2, compared with 78.5 for Gemini Decision 1 (Karpo).

What's new in the release?

Karpo positioned the release as a family rather than a single model. The six sizes share the same design principle: provide context, a question and options, and receive probabilities back. That consistency lets developers move between sizes without rewriting orchestration code.

  1. Six open-weight models ranging from 0.8B to 31B.
  2. An Apache-2.0 license on the Hugging Face repository.
  3. Support for up to 255 options per question.
  4. A 32,000-token context window with splitting for longer inputs.
  5. Multimodal inputs, including images and screenshots.

The breadth of the family matters for deployment. A 0.8B model can run on edge hardware; a 31B model targets high-end GPUs. Karpo said all sizes are available now, which lowers the barrier for teams that want to experiment with decision models before committing to a specific hardware footprint.

How does the model work technically?

Karpo Decision 31B is built on top of google/gemma-4-31B-it, a 31B instruction-tuned Gemma variant from Google. The fine-tuning teaches the model to map a question and a set of options to a probability distribution. Because the model does not generate tokens, inference cost is closer to a classification pass than to a language-model generation pass, which is the main reason latency stays near 94 ms on a single RTX PRO 6000 GPU with one request at a time (Hugging Face).

The typed question format is central to the design. A choice question asks the model to pick among alternatives; a noul question asks for a yes/no or true/false judgment; a score question asks for a rating. The model outputs calibrated probabilities rather than raw logits, which makes the outputs directly usable in downstream decision logic such as thresholds, argmax selection or ensemble averaging.

Karpo's blog lists a median latency of 95 ms, while the Hugging Face model card and the Karpo Decision page both report 94 ms per request on one GPU, one request at a time (Karpo; Hugging Face). The small difference is typical of benchmark measurement variance across test runs and hardware configurations.

How does it compare with Gemini Decision 1?

The most direct comparison in the release materials is on JevBench Capability, a benchmark for Jev-class open-weight models. Karpo Decision 31B scores 87.2, ranking No. 3 among 243 Jev-class open-weight models; Google's Gemini Decision 1 scores 78.5 (Karpo). The gap of 8.7 points is significant for a benchmark that measures capability on decision tasks, though JevBench covers a different slice of behavior than the Decision Index.

ModelDecision Index 0.4.1 FullJevBench CapabilityMedian latency on RTX PRO 6000
Karpo Decision 31B73.9 (No. 1, tied)87.294 ms
GLiDE 27B73.4 (No. 1, tied)——
Gemini Decision 1—78.5—

The comparison table combines figures published by Karpo and Hugging Face; neither GLiDE 27B nor Gemini Decision 1 has a published JevBench or latency figure in the release materials, which is why those cells are marked as not available.

What does the latency mean for real-world agents?

Latency of 94 ms per decision changes how agent frameworks are built. In a tool-selection loop that previously required a full generation — often 500 ms to several seconds — a decision model can score hundreds of candidate tools within a single forward pass. The 255-option limit covers large action spaces, and the 32k context window allows the model to incorporate system state and recent history before scoring.

The design also reduces cost. Token generation is the dominant cost driver for language models; eliminating it for the decision step means each decision costs a fraction of a generation. For enterprises running agent fleets at scale, the difference shows up in GPU utilization and per-request pricing.

What are the implications for the model market?

The release intensifies competition in the decision-model segment, where Google's Gemini Decision 1 had been the reference point. Karpo's open-weight approach under Apache-2.0 gives developers an alternative they can host on their own infrastructure, a key consideration for regulated industries. The tie at the top of the Decision Index also signals that open-weight models can match proprietary ones on private benchmark suites.

For Google, the result is notable: its own Gemma base model, fine-tuned by a third party, now outranks its proprietary Gemini Decision 1 on JevBench. That is a recurring pattern in the open-weight ecosystem, where base models become substrates for competitors' specialized fine-tunes.

What did Karpo say about the release?

In the announcement post on X, Karpo representative LeahW framed the release around speed and accuracy, citing the three headline numbers: the Decision Index score, the latency on RTX PRO 6000, and the JevBench comparison with Gemini Decision 1 (LeahW 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. Karpo Decision 31B 🏆 73.9 on Decision Index, tied for #1 on the Nov 12 board ⚡ 94 ms median per request on a single RTX PRO 6000 📊 87.2 on JevBench Capability vs. 78.5 for Google's Gemini Decision 1LeahW, Karpo representative, on X

What's next for Karpo and decision models?

Karpo said it has open-sourced all six model sizes, which suggests the company is betting on the open-weight distribution channel rather than a gated API. The next competitive questions are whether the Decision Index will be updated with new private tests, whether Google responds with a stronger Gemini Decision model, and whether downstream frameworks — such as agent SDKs and RAG pipelines — add native support for probability-scoring models like this one.

For buyers, the practical test is reproducibility: the 94 ms latency figure is measured on a single RTX PRO 6000 with one request at a time, and the Decision Index score comes from private tests. Teams evaluating Karpo Decision 31B will need to run their own workloads to see whether the calibrated probabilities hold up on their specific question types and context distributions.

Frequently asked

What is Karpo Decision 31B?

Karpo Decision 31B is an open-weight multimodal decision model fine-tuned from google/gemma-4-31B-it. It answers typed questions by returning a probability per option in one forward pass and is available under Apache-2.0 on Hugging Face.

How does Karpo Decision 31B compare with Gemini Decision 1?

On JevBench Capability, Karpo Decision 31B scores 87.2 versus 78.5 for Google's Gemini Decision 1, according to Karpo. On the Decision Index 0.4.1 Full board, it ties for No. 1 with a 73.9 score.

How fast is Karpo Decision 31B?

The model reports a median latency of 94 ms per request on a single RTX PRO 6000 GPU with one request at a time, per the Hugging Face model card. Karpo's blog lists a median latency of 95 ms.

Which model sizes did Karpo release?

Karpo released six open-weight models: 0.8B, 2B, 4B, 9B, 12B and 31B, all available on Hugging Face.

Sources

  1. Karpo — Decision Index Full 73.9, JevBench Capability 87.2 vs 78.5 for Gemini Decision 1, median latency 95 ms, six model sizes 0.8B-31B
  2. Hugging Face — Ranked No. 1 on Decision Index with Full score 73.9, median latency 94 ms, board assigns one rank to models within 0.9 points; karpo-decision-31B and GLiDE 27B share No. 1
  3. Karpo — 73.86 Decision Index 0.4.1 Full score, rank No. 1 (tied), 94 ms median latency, from Decision Index board of Nov. 11, 2026
  4. X (formerly Twitter) — Announcement of Karpo Decision family: 73.9 Decision Index, 94 ms latency, 87.2 JevBench vs 78.5 for Gemini Decision 1
  5. X (Twitter) — 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. Karpo Decision 31B 🏆 73.9…