SenseTime · Released, not on our provider
SenseNova Loop 9B (2026): what it is, how it scores against Qwen3.8 27B, and how to run it
SenseNova Loop 9B is SenseTime’s open-weight multimodal model, released on November 17, 2026, that runs 24 of its 32 layers twice to act like a 56-layer model with about 9B stored parameters, and SenseNova’s own charts show it matching Qwen3.8 27B on some tests but not all.
Use it now
Chat
This chat starts with Qwen3.8 Flash, not SenseNova Loop 9B.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- SenseTime
- Type
- Text
- Released
- 2026-11-17
- OpenRouter id
- None
Checked
SenseNova Loop 9B: facts, with sources
- The SenseNova-Loop GitHub page says SenseNova-Loop-9B was released on November 17, 2026, after SenseNova-Loop-9B-Preview in late October (the Preview’s Hugging Face card is dated October 30). AWQ and GGUF downloads were added on November 18.Source: github.com · checked 2026-11-18
- The code is published under the OpenSenseNova GitHub organisation, which describes itself as "SenseTime Foundation Model Sets"; the Hugging Face organisation is named sensenova.Source: github.com · checked 2026-11-18
- It has about 9B parameters (about 14.6B counting the repeated layers). The decoder stores 32 layers; layers 4 to 27 (24 layers) run twice with shared weights, so one forward pass makes 56 layer computations. It adds two-stream manifold-constrained hyper-connections (mHC) and alternates three linear-attention layers with one full-attention layer. It takes text and images and outputs text.Source: github.com · checked 2026-11-18
- The release supports a context of up to 512K tokens (the Preview was trained to 262,144). SenseNova says that with 4-bit weights, about 4.9 GB, and an 8-bit KV cache, inference at a 256K context needs roughly 15 GiB of GPU memory; at 64K, about 10 GiB.Source: github.com · checked 2026-11-18
- Weights are on Hugging Face (BF16, plus AWQ 4-bit with an MTP draft model, and GGUF in BF16, Q8_0, Q5_K_M and Q4_K_M) and ModelScope, under Apache 2.0. Hugging Face marks the main repository as gated with automatic approval. The README loads it through Transformers with trust_remote_code, and points vLLM and SGLang users to pull requests (vLLM #64832, SGLang #47281) rather than a released version; LightLLM has a Docker example.Source: huggingface.co · checked 2026-11-18
- SenseNova lists known limits itself: the model may miss requirements when given many complex tool definitions or constraints, performance can be inconsistent on coding and general agent tasks, and complex websites with 3D scenes and rich animation may need two to three rounds of refinement.Source: github.com · checked 2026-11-18
- The weights can be downloaded without a fee. We found no hosted API price for SenseNova-Loop-9B on the pages we read, and OpenRouter has no price because it does not list the model.Source: github.com · checked 2026-11-18
- SenseNova Loop 9B is not on OpenRouter’s model list (checked November 18, 2026), so this site cannot run it. This site is a text chat with other models, such as Qwen3.8 Flash and Inkling 2 Small.Source: openrouter.ai · checked 2026-11-18 · second-hand
Our table
SenseNova Loop 9B next to its Preview and Qwen3.8 27B
- SenseNova’s own chart (not tested by us) shows SenseNova-Loop-9B at 65.31 on AIDABench against 65.01 for Qwen3.8 27B, 72.34 against 69.82 on BIRD, 69.90 against 66.99 on GAIA, and level on AutomationBench at 36.67; it is behind on NL2Repo (41.36 against 44.60) and τ³-Banking (30.93 against 36.08).Source: github.com · checked 2026-11-18
Vendor benchmarks
SenseNova Loop 9B benchmark scores against the Preview and Qwen3.8 27B
People who search for SenseNova Loop 9B mostly want to know whether a 9B model can stand next to Qwen3.8 27B, the model SenseNova itself picks as its yardstick. By SenseNova’s numbers it does on some tests and not on others: it is level or ahead on four of the six shown, behind on two, and Terminal-Bench 2.1 has no Qwen3.8 27B bar. All of it is the maker’s own testing on agent, coding and data tasks, not an independent result. You cannot chat with this model here, but Qwen3.8 27B has its own table, and for small open-weight models you can chat with, see Inkling 2 Small and Hy4 Flash.
| Benchmark | SenseNova-Loop-9B | Preview | Qwen3.8 27B |
|---|---|---|---|
| AutomationBench | 36.67 | 14.33 | 36.67 |
| Terminal-Bench 2.1 | 59.55 | 40.45 | Not in the chart (Qwen3.6 27B: 63.40) |
| NL2Repo | 41.36 | 19.91 | 44.60 |
| AIDABench | 65.31 | 58.62 | 65.01 |
| BIRD | 72.34 | 71.38 | 69.82 |
| τ³-Banking | 30.93 | 18.56 | 36.08 |
| GAIA | 69.90 | 37.86 | 66.99 |
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to SenseNova Loop 9B you can use now
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
Inkling 2 Small
Thinking Machines Lab$0.30 / $1 (per million tokens)
- Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
- A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
- Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14
FAQ
SenseNova Loop 9B FAQ
What is SenseNova Loop 9B?
It is a multimodal model from SenseTime’s SenseNova team that reads text and images and writes text, built for reasoning, coding, tool use and long documents. It stores about 9B parameters and reuses a block of its layers twice, so one pass does the work of a deeper model. The weights were released on November 17, 2026.
What is a looped transformer?
A normal transformer passes data through each layer once. SenseNova Loop 9B sends the data through layers 4 to 27 twice, using the same weights both times, so 32 stored layers make 56 layer computations. SenseNova says this gives more depth without more stored parameters; it also adds two residual streams (mHC) to route information through the loop.
Does SenseNova Loop 9B match Qwen3.8 27B?
On some tests, by SenseNova’s own numbers. It scores 65.31 against 65.01 on AIDABench, 72.34 against 69.82 on BIRD and 69.90 against 66.99 on GAIA, and ties on AutomationBench at 36.67; it is behind on NL2Repo and τ³-Banking. These figures come from the maker’s charts and we did not run them. See Qwen3.8 27B for the other side.
How do I run SenseNova Loop 9B?
Download the weights from Hugging Face (sensenova/SenseNova-Loop-9B) and load them with Transformers using trust_remote_code. There are also an AWQ 4-bit version and GGUF files (Q4_K_M, Q5_K_M, Q8_0, BF16). SenseNova estimates about 15 GiB of GPU memory for a 256K context with 4-bit weights and an 8-bit KV cache. vLLM and SGLang support is still in pull requests.
Is SenseNova Loop 9B open source?
The weights are open under Apache 2.0, and SenseNova publishes code, with training scripts in separate repositories. Hugging Face marks the main repository as gated with automatic approval, so you accept access conditions before downloading. The technical report is not yet out for the Preview, and we did not find one for the release.
How is the release different from the Preview?
SenseNova says the release beats the Preview on all seven benchmarks it shows: AutomationBench rises from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55 and GAIA from 37.86 to 69.90. The context length also grows from 262,144 to 512K tokens. Both keep the looped design.
Can I chat with SenseNova Loop 9B here?
No. It is not on OpenRouter’s model list, so this site cannot run it. You can chat with Alibaba’s Qwen3.8 Flash instead: on this site a typical reply to Qwen3.8 Flash costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit.
Is this the official SenseNova website?
No. This page is unofficial and not affiliated with or endorsed by SenseTime or SenseNova. Facts link to the pages we read.