Commerce-1: open-weight 27B model scores agent decisions with zero generated tokens
yasintoy · reddit · 2026-10-10
infercrane released Commerce-1, a 27B open-weight model built for the action-selection step of commerce agents: it scores up to 255 candidate actions and returns typed decisions with probability distributions, with zero generated tokens.
- Built on perplexity-ai/pplx-decider-v1.1-27b with a noncausal full-attention decision backbone and dedicated readout
- Reproduced the full Decision Index 0.2.1 public suite: 60.99, across 150,317 eval rows
- Throughput of 8.836 decisions/sec on a single H200; three-decision fixture p50 340ms
- Weights on Hugging Face, runtime Apache-2.0; BF16 weights 54 GiB, qualified only on one H200
- The model only recommends actions — inventory truth, payments and fraud rules stay in application code
More from coding & agent
- Garry Tan: agents auto-improved eval loops for 16 hours with zero human input — garrytan · 2026-10-11
- NVIDIA 推出 Jetson Agent Skills，让编码 Agent 懂边缘硬件 — blaizedsouza · 2026-10-11
- Agent auto-fixing JS errors uncovers WebKit WebSocket payload truncation bug — Vjeux · 2026-10-11
- SkyPilot meetup recap: GPU scheduling pain, Holo4 training on K8s, and SGLang's roadmap — skypilot_org · 2026-10-11
- ODS V3 pre-release turns any PC, Mac or Linux box into a private AI server — tom_doerr · 2026-10-10
- Dev praises Hugging Face docs as the best resource for learning LLMs — ayushthakur0 · 2026-10-10