SenseNova-Loop-9B released with looped architecture and 512K context support

·2026.11.17 23:08

Key point

The model extends 32 physical layers to 56 executions via looped computation and supports 512K token context.

Details

  • SenseNova-Loop-9B officially released, featuring a looped architecture that reuses layer weights to expand computational depth within a 9B parameter budget.
  • Competitive performance with Qwen3.8-27B in agentic and data analysis tasks, scoring 41.36 versus 44.60 on NL2Repo and 65.31 versus 65.01 on AIDABench.
  • Architecture uses 24 decoder layers executed twice to produce 56 layer computations, combined with two-stream manifold-constrained hyper-connections (mHC) and hybrid attention.
  • Supports multimodal input (text and images) with a 512K token context length, requiring approximately 15 GiB of GPU memory for 256K context inference with 4-bit weights and 8-bit KV cache.
  • Official AWQ 4-bit quantized version available on launch day for easier deployment.

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.