Cambricon announced Day 0 adaptation of SenseTime's newly released SenseNova Loop-9B, bringing the model's inference pipeline onto its MLU platform via vllm_mlu. The work covers static unrolling of the Loop layers, reuse of existing Qwen3-Next MLU operators, and Triton-fused optimization of mHC routing, enabling low-memory deployment of this loop-computing architecture on domestic accelerators.
Termsslug: recurrent-compute
Recurrent Compute
循环计算
Increasing reasoning depth by iterating the same network rather than emitting intermediate tokens.
Curated Intelligence
IT之家 · 智能时代T2·Infra
58 pts
新智元 (微信公众号)T2·Models
62 pts
Researchers from Lehigh University, UBC and the Vector Institute propose Lev, a decision model that scales reasoning depth through recurrent loops rather than generated reasoning tokens, keeping simple requests at tens of milliseconds while giving hard ones more compute. It reports up to 13% accuracy gains on decision tasks, with code, models and datasets released openly.