SenseNova Releases Loop-9B Preview: Looped Decoder Hits 56-Layer Depth from 32 Layers, 256K Context
jacek2023 · reddit · 2026-10-30
SenseNova open-sourced SenseNova-Loop-9B-Preview on Hugging Face, a multimodal model for long-context and agentic workflows spanning document understanding, coding, and tool use.
Architecture highlights:
- Looped decoder: 24 layers executed twice with shared weights, yielding 56-layer execution depth from 32 distinct layers without duplicating parameters
- Two-stream manifold-constrained hyper-connections (mHC): learned mappings route information across two residual streams in the looped section
- Hybrid attention: linear and full attention interleaved in a repeating four-layer pattern
- Training on text, images, code, and tool trajectories at sequence lengths up to 256K tokens
The looped-reuse design is a parameter-efficient bet on buying execution depth; official deployment configs are recommended for best performance.
More from Models
- OpenAI's subscription pain: low limits and unbanked resets drag utilization — teortaxesTex · 2026-10-30
- Subscription utilization test: Claude Max hits 80-90%, OpenAI only ~50% — teortaxesTex · 2026-10-30
- Unsloth rebuts GSQ-RCO: DeepSeek-V4.1-Flash 1bit AIME25 is 91.88%, not 100% — danielhanchen · 2026-10-30
- SAM3 now runs offline on your GPU: ONNX export hits 0.37s on M5, 0.08s on RTX 3080 Ti — CSProfKGD · 2026-10-30
- Same model, different hardware: dev finds A100s vs B300s produce different outputs — Kyrannio · 2026-10-30
- CMU paper: one algorithm for pretraining and post-training — unified training beats staged pipelines — heghbalz · 2026-10-30