Karpo Decision 31B: Open-Weight Model Outputs Decision Probabilities in One Forward Pass
alifcoder · x · 2026-11-17
Karpo released Karpo Decision, a family of six open-weight models (0.8B–31B) built for agents. The core idea is "a readout, not generation": given a state and typed questions with options, a linear readout plus softmax returns a probability per option in a single forward pass, with no token decoding—so answers are always among the options.
Key numbers:
- 73.9 on Decision Index, tied #1 on the Nov 12 board
- 94 ms median latency on a single RTX PRO 6000
- 4.8 points of calibration error; up to 255 options per question
- 87.2 on JevBench Capability vs. 78.5 for Google's Gemini Decision 1; 90.5 vs 79.7 on New Yorker caption
Three question types (choice / noul / score) are supported, using branch attention where each branch reads only the shared state block and itself. Weights take 62.5 GB on one GPU. Suggested first uses: search, recommendations, tool choice, model routing, and browser actions.
More from coding & agent
- Matt Pocock's /constrain Idea: Cage Agents With Brutal Lint Rules — mattpocockuk · 2026-11-17
- Editing WordPress via MCP Feels Like "Superpowers"; Contributors Want It in Core — VeryWellVersed · 2026-11-17
- Self-hosted MCP endpoint exposes 38 database engines, read-only, creds stay server-side — Green_Competition_21 · 2026-11-17
- Stop Over-Trusting AI Coding Leaderboards: CACM Paper Cites OpenAI's SWE-bench Verified Reversal — EhudReiter · 2026-11-17
- SubCast: open-source bilingual podcast transcripts with per-sentence AI Q&A — vista8 · 2026-11-17
- "English was always meant to be the final programming language" — tlakomy · 2026-11-17