Karpo Decision 31B open-weights model tops Decision Index with one-forward-pass answers
nikola_mr64990 · x · 2026-11-21
Karpo released a family of six open-weight decision models (0.8B–31B) with a fundamentally different architecture: given a state, questions, and typed options, a linear readout plus softmax scores every option in a single forward pass — no token generation, answers always restricted to the options, up to 255 per question.
Key results:
- 73.86 on Decision Index 0.4.1, tied for #1 on the Nov 12 board
- 94 ms median latency on a single RTX PRO 6000, one request at a time
- 90.0% on Mind2Web element-picking vs Gemini Decision 1's 86.2%; 87.2 on JevBench Capability vs 78.5
- 4.8 points calibration error; multiple questions answered in one forward pass via branch attention
Suggested use cases: search, recommendations, tool choice, model routing, and browser-agent actions — point it at the slowest decision in your stack. Weights and inference code are on Hugging Face.
More from Models
- Leaked Gemini App screenshot exposes internal reasoning and system prompt with teen safety rules — Bigwaliwigi · 2026-11-21
- NandiBERT-150M and 450M weights now live on Hugging Face under Apache 2.0 — kalyan_kpl · 2026-11-21
- OpenAI researcher hints at mystery model "Inkling" in babysitting tweet — yiwenyuan98 · 2026-11-21
- LlamaIndex benchmark: Gemini 4 Neon leads tables and visual grounding at 6.2¢/page — llama_index · 2026-11-21
- Inkling 2 Mini quantized to 13.9GB runs at ~60 t/s decode on a single RTX 4090 — indrasmirror · 2026-11-21
- Claude Code reportedly beat GTA VI solo in 12 hours, digging up 500+ hidden secrets — nikola_mr64990 · 2026-11-21