9.66B stored / approximately 14.6B inference parameters (dense shared-weight looped model; exact checkpoint inventory: 9,656,250,480 parameters)
262,144 tokens
Apache-2.0
Preview open-weight multimodal language model for long-context document understanding, coding, tool use, and agentic workflows. It accepts text and images and generates text, with optional thinking controlled through its chat template. Instead of increasing stored parameter count conventionally, SenseNova Loop reuses a 24-layer decoder block twice: layers 0-3 execute once, layers 4-27 execute twice with shared weights, and layers 28-31 execute once. This produces 56 decoder-layer executions from 32 distinct stored layers, which SenseNova describes as approximately 14.6B inference parameters from a roughly 9B backbone. The repeated section uses two-stream manifold-constrained hyper-connections, with learned read, stream-mixing, and residual-write mappings; Sinkhorn normalization constrains the mixing map using 20 iterations. The 4,096-dimensional decoder combines three linear-attention layers followed by one full-attention layer in each repeating four-layer group. Its configuration specifies 16 query heads, four key/value heads, a 12,288-dimensional feed-forward layer, partial RoPE, a 248,320-token vocabulary, and a 27-layer, 1,152-dimensional Qwen3.5-style vision encoder. The checkpoint includes one Multi-Token Prediction layer, and MTP acceleration is available; DSpark and DFlash2 draft models are still being trained. Documented training combines long sequences, multimodal inputs, code, and supervised tool trajectories at lengths up to 262,144 tokens, but the complete pretraining corpus and post-training recipe are not disclosed. Publisher-reported results include 71.38 on BIRD-Dev, 66.89 on WorkBuddy Bench, 58.87 on AIDABench, 48.89 on TeamBench, 40.45 on Terminal Bench 2.1, 18.56 on tau3-Banking, 13.72 on AutomationBench, and 70.70 on the internally developed RaccoonPPTBench. The comparison chart shows particularly strong parameter efficiency but mixed results against larger Qwen and Nanbeige models; evaluation prompts, attempt aggregation, and several harness details are not documented in the release card. Transformers inference requires version 5.18.0 and trust_remote_code. LightLLM support is documented with a four-GPU example, while SGLang support is provided through an active integration pull request. SenseNova also publishes TorchTitan recipes for SFT, DPO, and MTP training, plus SenseSpec recipes for DSpark and DFlash2. The release contains eight main weight shards and a separate MTP file totaling approximately 38.14 GB. Although the README describes a BF16 backbone with FP32 mHC routing parameters and the configuration declares BF16, Hugging Face's shipped safetensors metadata reports approximately 9.41B F32 parameters and 243M BF16 parameters, so the currently downloadable artifact is predominantly FP32. The technical report is marked TBD.