MachinePulse open-sources Karpo-Decision-31B, tops Decision Index at 73.9
Negative-Whereas3307 · reddit · 2026-11-16
MachinePulse released Karpo-Decision-31B, an Apache 2.0 multimodal decision model fine-tuned from Gemma 4 31B. It scored 73.9 on the Nov 12 Decision Index — the top Full score among 300 entries, tied with GLiDE 27B — while claiming 2x the speed of Jev at 20% lower cost.
How it works
- A trained linear readout scores supplied answer options, supporting choice (up to 255 options), yes/no and score questions, outputting probabilities directly
- Branch attention answers multiple questions in one forward pass; handles text and images including web screenshots, with a 32,768-token sequence limit
- Notable training detail: full fine-tuning pushed the public score to 75.09 but cut unfamiliar-domain accuracy by 2.3 points; averaging five checkpoints (three full-FT + two LoRA) recovered 1.6 points with no extra inference cost
Results: reported wins over Gemini Decision 1 (When2Call 90.4% vs 79.9%, Mind2Web 90.0% vs 86.2%, etc.); 95ms median latency on one RTX PRO 6000. Weights and inference code are on Hugging Face (62.5GB BF16).
Related event: MachinePulse open-sources Karpo decision models, topping Decision Index(2 posts)→
More from Models
- Nex-N2.6, a GLM-5-Next-based multimodal model, trends on Hugging Face — nex-agi · 2026-11-16
- New player Maxwell launches V1 Flash: 35B/A3B MoE, 256K context, Apache 2.0 — AdinaYakup · 2026-11-16
- FLUX-mimic-x: 7B video-action model tops RoboLab leaderboard at 51.4%, zero-shot across embodiments — robrombach · 2026-11-16
- Qwen4-max launch page called out: no model card, no benchmarks, no pricing — petrusenko_max · 2026-11-16
- SLDBench v2: Claude Opus 5.5 beats classic fits using 4% of compute budget — jiqizhixin · 2026-11-16
- Grok Bot 1.0 lands with bots that join meetings, take initiative and self-improve — mattyp · 2026-11-16