Karpo Decision 31B open weights: a decision model scoring all options in one forward pass
_jaydeepkarale · x · 2026-11-17
Karpo released Karpo Decision, a family of six open-weight decision models (0.8B–31B) built for agents.
Instead of generating tokens, the model uses a linear readout + softmax: given a state and typed questions (choice with up to 255 options, true/false, or score), it outputs a probability for every option in a single forward pass. Multiple questions are encoded in one sequence with branch attention, so each branch reads only the shared state and itself.
Key numbers:
- 73.9 on the Decision Index, tied for #1 among 300 models (Nov 12 board)
- 94 ms median latency per request on a single RTX PRO 6000
- When2Call 90.4% vs 79.9% for Gemini Decision 1; 87.2 on JevBench Capability vs 78.5
Suggested first uses: search, recommendations, tool choice, model routing, and browser actions — pointing it at the slowest decision in your stack. Weights are free to download on Hugging Face.
More from coding & agent
- Swamp compiles repeated skills into deterministic workflows: 8x fewer tokens, 2x faster — aigclink · 2026-11-17
- Riley Brown: After 50+ DMs, Nothing Beats Claude Code Projects Right Now — angrypenguinPNG · 2026-11-17
- MCP server turns your saved reels and posts into searchable notes for Claude — PaRtHuS · 2026-11-17
- Stardock CEO shows agents building games across PCs and Macs, with hard-won engineering lessons — draginol · 2026-11-17
- sandbox-mcp: Hanselman's MCP server gives AI agents disposable real Windows 11 machines — unixterminal · 2026-11-17
- ProceduralTerrains: open-source Three.js GPU terrain generator hits 2.0 — anselm · 2026-11-17