karpo-decision-31B
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.8/10
Weights public since 5 Nov 2026. From 12 Nov the card's license is MachinePulse Community: commercial use under US$10M revenue, with attribution for products offered to third parties. Card, Python API and a /v1/systemone server. No hosted API and no SLA.
- Capability7.6/10
Choice, noul and score, text only, one pass. DI 0.4.1 file of 19 Nov 2026: Full 73.86, strict #1 of 326. REAX 1.3 is 72.64. Board ECE 0.048; the card names no method. Median 94 ms, p95 586 ms.
- Adoption2/100
26 Hugging Face downloads and 3 likes on 12 Nov 2026. The inference repo is public. No hosted API, framework listing or named production user.
Vendor claims
- Ranked #1 (tied) on Decision Index 0.4.1, Full 73.86, in a group with GLiDE 27B and REAX 1.3 because the board UI groups models within 0.9 points and can hide inference techniques.[Vendor claim — not independently verified]
News covering this model
Read 12 Nov 2026 from the model card, the inference repo and
data/index.jsonon the Decision Index space (generated 11 Nov 2026, 14:28 UTC, 297 entries).
karpo-decision-31B is a text decision model from MachinePulse, fine-tuned from google/gemma-4-31B-it with a linear readout over option codes. A state plus named questions comes back as a probability for every option, in one forward pass, with no generated text. The card lists choice (up to 255 options), noul and score. Context is 32,768 tokens per sequence; a longer request is split by question, each part keeping the full state. Vision weights are in the checkpoint and unused. Loading it as a chat model does not reproduce the answers.
The weights were first published on 5 Nov 2026. The card says that from 12 Nov 2026 the files in the repo, including those first published on 5 Nov, are under the MachinePulse Community License. Commercial use is allowed. Products and services offered to third parties must show an attribution line. Production commercial use by an entity whose annual revenue, with affiliates, exceeds US$10 million needs a separate agreement. Copies taken before 12 Nov under Apache-2.0 keep the rights that licence granted. The inference code at machinepulse-ai/karpo-decision stays Apache-2.0. There is no hosted API.
Decision Index 0.4.1 (11 Nov 2026)
Sorted by Full score (balanced_skill) in the space file data/index.json, generated 2026-11-11T14:28:34Z, 297 entries. The same file still has Jev 1.13 at Full 59.22. This is the same edition the catalog already uses. It is not mixed with Artificial Analysis or JevBench.
| Rank | Model | Full |
|---|---|---|
| 1 | REAX 1.3 | 73.98 |
| 2 | karpo-decision-31B | 73.86 |
| 3 | GLiDE 27B | 73.37 |
karpo’s own row: public index 75.05, ECE 0.0477, median 94.3 ms, p95 585.9 ms. The catalog latency fields are that median and that p95, not a published p50/p99 pair. The card’s public-index figure matches the board. The card’s “calibration error 4.8 points” is that ECE, rounded.
Author claim (UnverifiedClaim). The card says the model is #1, tied, because the board UI gives one rank to models within 0.9 points of the leader and the chart they show hides inference-technique rows. On that reading, GLiDE 27B, karpo and REAX 1.3 share first. The 11 Nov file does not tie them: 73.98, 73.86 and 73.37 are three scores. The catalog rank is the strict sort. REAX 1.3 stays a news item, not a catalog page.
Decision Index file of 12 Nov 2026
data/index.json generated 2026-11-12T12:34:44Z, 300 entries. karpo’s Full is still 73.86 (public 75.05, ECE 0.0477, median 94.3 ms, p95 585.9 ms). The strict order at the top is now karpo, GLiDE 27B at 73.37, then REAX 1.3 at 72.64. Jev 1.13 is still 59.22. The 0.49-point gap to GLiDE is not the 0.9-point UI band, and this page still does not call it a tie.
- Capability 7.6 = types 10 × 0.30 + calibration 4 × 0.25 + latency 7 × 0.25 + accuracy 9 × 0.20 = 3.00 + 1.00 + 1.75 + 1.80 = 7.55 → 7.6. Types, calibration and latency are unchanged. Accuracy: half-up(8 × 73.86 ÷ 59.22) = 10, capped at 8, then 9 because this file makes karpo an untied first. There is still no second independent board for karpo. Maturity 4.8 and adoption 2 stay.
Decision Index file of 19 Nov 2026
data/index.json generated 2026-11-19T14:42:56Z, 326 entries. karpo’s Full is still 73.86, still the strict first. GLiDE 27B is still 73.37. REAX 1.3 is still 72.64. Jev 1.13 is still 59.22. The row’s ECE, median and p95 are unchanged. 4.8 / 7.6 / 2 stays: it is still an untied first, and there is still no second board.
Scores
Weights in the quadrant doc are unchanged. Reference for the proportional rule remains Jev 1.13 Full 59.22.
- Maturity 4.8 = availability 7 × 0.30 + docs 7 × 0.25 + integrations 3 × 0.25 + support 1 × 0.20 = 2.10 + 1.75 + 0.75 + 0.20 = 4.80. Availability 7: public weights, with a revenue cap, not an unrestricted Apache grant. Docs 7: card, Python API and HTTP server. Integrations 3: its own
/v1/systemoneserver, not a framework listing. Support 1: a licence contact address, no SLA. - Capability 7.6 (was 7.4 on the 11 Nov file) = types 10 × 0.30 + calibration 4 × 0.25 + latency 7 × 0.25 + accuracy 9 × 0.20 = 3.00 + 1.00 + 1.75 + 1.80 = 7.55 → 7.6. Types 10: choice, noul and score. Calibration 4: the board publishes an ECE and the card does not name a method (same treatment as Jeffan). Latency 7: median 94 ms is inside the ≤ 0.25 s band; p95 is 586 ms, so this is not the ≤ 100 ms and ≤ 500 ms band. Accuracy: half-up(8 × 73.86 ÷ 59.22) = 10, capped at 8, then 9 because the 12 Nov file makes this an untied first. The 11 Nov file had it second, behind REAX at 73.98, and the cap of 8 stood.
- Adoption 2 = engagement 2 × 0.35 + ecosystem 4 × 0.35 + production 0 × 0.30 = 0.70 + 1.40 + 0 = 2.10 → 2.
