MachinePulse Releases Karpo-Decision-31B, an Open Multimodal Model Ranking #1 on Decision Index

·2026.11.17 00:04

Key point

The Gemma 4 31B fine-tune achieves the highest Full score among 300 entries and runs 2x faster than Jev at 20% lower cost.

Details

  • MachinePulse released Karpo-Decision-31B, a Gemma 4 31B fine-tune for structured decisions over text and images, scoring 73.9 on the Nov 12 Decision Index board.
  • The model shares the #1 rank with GLiDE 27B under the board’s tie rule, with a Full score weighted 60% for familiar skills and 40% for new domains.
  • Performance metrics include 70.7 on private familiar skills and 78.6 on private new domains, while the public index score of 75.0 is excluded from the Full score.
  • Reported median latency is approximately 95 ms per request on one RTX PRO 6000, with BF16 weights requiring approximately 62.5 GB of memory.
  • The model uses a trained linear readout and branch attention to handle up to 255 options per choice question in a single forward pass.
  • Weights and inference code are released under Apache 2.0, with a documented sequence limit of 32,768 tokens.
  • Benchmarks show Karpo-31B outperforming Google’s Gemini Decision 1 on When2Call (90.4% vs 79.9%) and Mind2Web (90.0% vs 86.2%).

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.