model releaseSensetime

SenseNova-Loop-9B released: 9B parameters, 56 layer executions, 512K context

TL;DR

SenseNova has released SenseNova-Loop-9B, a roughly 9B-parameter multimodal model that runs 56 decoder-layer executions per forward pass from 32 distinct layers. It supports up to 512K tokens of context and, per the model card, needs about 15 GiB of GPU memory at 256K context with 4-bit weights and an 8-bit KV cache.

3 min read
0

SenseNova has released SenseNova-Loop-9B, a multimodal model with approximately 9B total parameters that executes 56 decoder-layer computations per forward pass while storing only 32 distinct layers. It accepts text and images, generates text, and supports a context length of up to 512K tokens. Weights and code are available on Hugging Face, where access requires accepting the repository's conditions. The release follows SenseNova-Loop-9B-Preview, published on 2026-10-31.

Architecture

The model reuses layer weights through a looped design. Layers 0 to 3 run once, layers 4 to 27 (a 24-layer block, range [4, 28)) run twice in sequence with shared weights, and layers 28 to 31 and the output head finish the pass. The model card lists approximately 14.6B inference-equivalent parameters.

Key specifications:

  • Hidden dimension: 4,096
  • Feed-forward dimension: 12,288
  • Vocabulary size: 248,320
  • Attention: a repeating four-layer pattern of three linear-attention layers followed by one full-attention layer; full attention uses 16 heads and 4 KV heads
  • Two-stream mHC: manifold-constrained hyper-connections with 2 residual streams, with the stream-mixing map constrained by Sinkhorn normalization (20 iterations)
  • Precision: BF16 backbone, FP32 mHC routing parameters

The card states that a compatible inference implementation must load the loop range, pass count and mHC routing parameters and preserve the same computation order. Training covered text, images, code and tool-use trajectories, with context extended to 512K tokens.

Benchmarks (reported by SenseNova)

All figures below come from SenseNova's own evaluations. The company says the release outperforms the Preview on all seven benchmarks shown.

Benchmark Preview Release
AutomationBench 14.33 36.67
Terminal-Bench 2.1 40.45 59.55
GAIA 37.86 69.90
BIRD 71.38 72.34

Against Qwen3.8-27B, SenseNova reports 41.36 versus 44.60 on NL2Repo, 65.31 versus 65.01 on AIDABench, and a match at 36.67 on AutomationBench. The company describes the model as competitive with Qwen3.8-27B on agentic and data analysis evaluations.

Loop scaling

SenseNova tested one to four loops, holding training data and other settings constant, with a single pass (Loop 1) of the original 9B model as the baseline. AIDABench-QA and Workbuddy-office peak at three loops, improving 6.42% and 16.93% over baseline. LiveCodeBench V6 and AIDABench data visualization peak at four loops, improving 8.10% and 8.55%. Moving from two to four loops adds 1.04 percentage points on LiveCodeBench V6 and 1.35 on AIDABench data visualization. Both the Preview and the release default to two loops to balance inference efficiency. The company says it plans to explore Dynamic Loop architectures that adapt loop count to each input.

Deployment

With 4-bit quantization the weights occupy approximately 4.9 GB. An official AWQ 4-bit version is available at launch. Estimated minimum GPU memory:

Context 8-bit KV cache 4-bit KV cache
64K ~10 GiB ~9 GiB
128K ~12 GiB ~10 GiB
256K ~15 GiB ~12 GiB

Hosted API pricing is not yet disclosed.

What this means

The model trades repeated computation for stored parameters: it keeps 32 distinct layers in memory but performs 56 layer executions, which SenseNova positions as a way to add depth within a compact parameter budget. The reported memory figures (about 15 GiB at 256K context with 4-bit weights and an 8-bit KV cache) place long-context multimodal inference within a single-GPU memory range for this model. The loop-scaling results show gains beyond the default two loops on several tasks, so the default setting is an efficiency choice rather than a ceiling. The benchmark claims are SenseNova's own, and the Qwen3.8-27B comparison is mixed: ahead on AIDABench, level on AutomationBench, behind on NL2Repo. Because the checkpoint depends on a specific loop and mHC execution order, deployment requires an inference stack that implements that configuration.

Related Articles

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use

Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.

model release

Tencent releases Hy4 Flash Base: 220B-parameter MoE, 12B active, 1M-token context

Tencent has published Hy4 Flash Base on Hugging Face: a 220B-parameter mixture-of-experts model with 12B active parameters per token, text and image input, and a 1M-token context. Tencent calls it its most cost-effective Hy model to date.

Comments

Loading...