Model Info
| Field | Value |
|---|---|
| Model ID | sensenova/SenseNova-Loop-9B |
| Vendor | SenseTime |
| Parameters | ~9B stored parameters; ~14.6B inference-equivalent compute (of 32 distinct decoder layers, layers 4-27 run twice — 56 total layer executions from only 32 stored weight sets) |
| Context Window | 512,000 tokens (per the official Hugging Face model card; double the Preview's 262,144 tokens) |
| Input pricing (USD/1M tokens) | Not available (open weights only, no official hosted API) |
| Output pricing (USD/1M tokens) | Not available (same as above) |
| Open source | Yes (Apache License 2.0) |
| Release date | 2026-11-17 |
| Official announcement | SenseNova-Loop-9B model card (Hugging Face) |
| Hugging Face | sensenova/SenseNova-Loop-9B |
| Family | SenseNova Loop (the stable release, 18 days after the Preview opened-sourced on 2026-10-30) |
Highlights
- GAIA jumps from 37.86 to 69.90: the official Hugging Face model card's "From Preview to Release" benchmark chart shows the stable release beating the Preview on all seven charted benchmarks, with GAIA posting the largest gain (+32.04 points, nearly doubling)
- Terminal-Bench 2.1 climbs from 40.45 to 59.55, AutomationBench from 14.33 to 36.67 — the latter exactly ties Qwen3.8-27B, a model with roughly 3x the parameters (also 36.67)
- Context window doubles from 262,144 to 512,000 tokens, while the looped architecture itself is unchanged (32 distinct decoder layers executed 56 times, with two-stream mHC)
- Launch day ships official AWQ 4-bit and GGUF weights: the 4-bit quantized weights come in at roughly 4.9GB, and 256K-context multimodal inference is estimated to need only about 15GiB of GPU memory with an 8-bit KV cache
- Day-0 Cambricon MLU support: Cambricon's
vllm_mlustatically unrolls the looped architecture into 56 sequential execution slots, reusing vLLM's existing scheduling and KV-cache machinery — though Pandaily's coverage notes these throughput claims haven't been independently verified
Benchmarks
Figures from SenseTime's official Hugging Face model card. Scale: baseline y=300, plot top y=60, domain 0-80 points, k=(300-60)/80=3 units per point, bar top y=300-score*3.
| Benchmark | Score (stable) | Predecessor (Preview) | Best competitor (Qwen3.8-27B) |
|---|---|---|---|
| GAIA | 69.90 | 37.86 | Not available |
| Terminal-Bench 2.1 | 59.55 | 40.45 | Not available |
| AutomationBench | 36.67 | 14.33 | 36.67 (tied) |
| AIDABench | 65.31 | Not available | 65.01 |
| NL2Repo | 41.36 | Not available | 44.60 (trails) |
| BIRD | 72.34 | 71.38 | Not available |
All figures are SenseTime's own results from the official Hugging Face model card, not yet independently reproduced. Qwen3.8-27B's parameter count is roughly double the stable release's inference-equivalent compute (~14.6B).
vs. Predecessor and Competitors
Against the Preview, the stable release's biggest gains are on GAIA (+32.04 points, nearly doubling) and Terminal-Bench 2.1 (+19.1 points) — both benchmarks that demand multi-step tool chains and complex task planning, which suggests the 18-day iteration mainly strengthened agentic ability rather than just padding scores. AutomationBench is the most concrete "small model beats big model" evidence here: the Preview trailed the (differently-sized) Qwen3.5-27B at 10.2 points behind, and the stable release now exactly ties Qwen3.8-27B at 36.67.
Against Qwen3.8-27B, AutomationBench ties and AIDABench edges ahead slightly (65.31 vs 65.01, within noise), but NL2Repo trails by 3.24 points (41.36 vs 44.60) — code-generation-adjacent tasks remain a weak spot, consistent with the Preview-era pattern of falling behind larger models on Terminal Bench and financial-workflow tasks.
Pricing hasn't changed: the stable release, like the Preview, ships as open weights only with no official API, so it isn't competing on price. What's new is the official AWQ 4-bit quantization and day-0 Cambricon support, which lower the cost of self-hosting rather than the cost of calling an API — 4.9GB of quantized weights plus roughly 15GiB of GPU memory (at 256K context) makes consumer-grade hardware a more realistic deployment target than before.
What This Means for Agent Development
Doubling the context window (262K→512K) alongside the sharp jump in agentic benchmarks puts the stable Loop-9B closer to being something you'd actually run an agent on, rather than just a showcase of architectural efficiency.
- If you're running agents on consumer or edge GPUs: the official AWQ 4-bit quantization (~4.9GB) combined with roughly 15GiB of GPU memory for 256K-context inference gives you an officially-validated quantization path, which is less guesswork than the Preview era.
- If you're building long-context, multi-step tasks (document analysis, long-running conversational agents): the 512K context paired with the GAIA and Terminal-Bench 2.1 gains is worth evaluating — but SenseTime's own "Known Limitations" section flags inconsistent instruction-following with many complex tool definitions and multiple constraints, plus uneven performance on long-horizon coding and agent tasks, so validate on a small scope first.
- Not a fit: scenarios needing a stable public API or SLA guarantees — there's still only open weights, no official hosted service. Don't treat Cambricon's MLU throughput numbers as capacity-planning input either; Pandaily's coverage is explicit that those are the vendor's own unverified claims.
Today's Takeaway
Only 18 days separate the Preview from the stable release, yet GAIA nearly doubled and AutomationBench went from trailing to tying a ~27B-class competitor — a sign that "ship a Preview to gather community feedback, then iterate fast into a stable release" is becoming a standard cadence among Chinese open model labs, not a SenseTime one-off (Tencent's Hy4 flash and MiniMax's recent releases follow the same pattern). But what stands out more than the benchmark gains is the discipline on display: SenseTime's own model card plainly lists where long-horizon tasks still wobble, and Pandaily explicitly flags that Cambricon's throughput claims are unverified — a release that moves fast while drawing its own verification boundaries clearly.
References
- SenseNova-Loop-9B model card (Hugging Face, stable release)
- SenseTime Releases Final SenseNova-Loop-9B With Cambricon Day-0 Support (Pandaily, 2026-11-17)
- IT Home: SenseNova-Loop-9B stable release coverage (in Mandarin)
- IT Home: Cambricon MLU adaptation coverage (in Mandarin)
- SenseNova-Loop-9B-Preview model card (this site, 2026-10-31)