Loop-9B: 9B Model Matches 27B on Agent Tasks at Half Speed on 16GB Mac
SenseNova Loop-9B uses a looped transformer architecture to run 56 layer passes with only 9B parameters, achieving 27B-level performance on agent and coding tasks while fitting in 10GB VRAM, but generation speed is halved compared to standard 9B models.
Three Key Takeaways
This approach works. SenseTime released SenseNova Loop-9B on November 17: a 9B-parameter model that runs the same 24-layer block twice, delivering 56 layer computations (4 + 24×2 + 4) and matching 27B-scale agent/data tasks within 10GB VRAM; 4-bit weights are only 4.9GB.
Your 16GB Mac can run it now. The community has ported Loop-9B to Apple's MLX framework; standard mlx-lm loads it directly. The 4-bit version weighs 5.1GB and runs at ~22 tok/s on an M5 Max.
There is no free lunch. Looping means each token is computed twice, so generation speed is roughly halved versus a regular 9B model (community measured 49 vs 89 tok/s). It trades time for memory, not speed for quality.
The real question is not "can it run" but: are you willing to accept half the speed to get a model that fits in 10GB VRAM and still handles 27B-level agent workloads?
1. What SenseTime Actually Released
Timeline:
October 30: Open-sourced SenseNova Loop-9B-Preview (Apache 2.0), benchmarked against Qwen3.5-27B.
November 17: Released the official version, optimized for agent, code, and data analysis; benchmark target updated to Qwen3.8-27B; Cambricon adaptation completed same day.
Both share the same core: Looped Transformer . Traditional models are a one-way pipeline; Loop-9B folds the middle into a boomerang:
Architecture Blocks
Entrance (layers 0–3, 4 layers): Project input into hidden space.
Shared Loop Block (layers 4–27, 24 layers): Weights stored once; data passes through, loops back, passes again.
Exit (layers 28–31, 4 layers): Decode final token.
Result: physically 32 layers, but 56 layer computations per forward pass (4 + 24×2 + 4). Equivalent activated parameters ~14.6B, yet VRAM cost stays at 9B.
Two key innovations:
mHC (manifold-constrained hyper-connection): Borrowed from DeepSeek research; adds two residual streams inside the loop block, uses 20 Sinkhorn-Knopp iterations to keep the mixing matrix on a doubly-stochastic manifold, preventing signal explosion/collapse. Official ablation: +4.42 GPQA Diamond, +13.34 TeamBench.
Mixed Attention: 3 linear attention layers + 1 full attention (Gated DeltaNet style) per 4-layer group, reducing long-context compute.
In one sentence: Others stack parameters; this model stacks "thinking passes."
2. Where the 27B Claim Holds: Official Benchmarks (Self-Reported, Not Independently Verified)
All scores below come from SenseTime's own release; no third-party replication seen yet.
Official Benchmark Comparison
NL2Repo (code repo generation): Loop-9B 41.36 vs Qwen3.8-27B 44.60 — slightly lower.
AIDABench (agent/data analysis): Loop-9B 65.31 vs Qwen3.8-27B 65.01 — slightly higher.
Preview version showed stronger gaps against Qwen3.5-27B: WorkBuddy 66.89 vs 60.53, RaccoonPPT 70.70 vs 59.08, AutomationBench 13.72 vs 10.17.
Interpretation:
On agent, office automation, and data analysis tasks, 9B reaches the 27B threshold, sometimes exceeding it — source of the "9B runs like 27B" claim.
On benchmarks testing hard knowledge and professional depth (Terminal Bench 2.1, T³-Banking), 9B does not overturn 27B's bulk advantage.
Critics noted Preview compared against older Qwen3.5-27B; official version switches to Qwen3.8-27B.
Conclusion: It matches 27B on specific tasks (agent/code/office), not across the board. Don't be misled by "9B beats 27B" headlines.
3. Your 16GB Mac: What It Can Actually Run
Good news: the Mac port is ready. Team netprojects ported Loop-9B to MLX (Apple's official ML framework); standard mlx-lm loads it without patches. They built three quantizations and cross-checked against SenseTime's original PyTorch code:
MLX Quantization Versions (tested on M5 Max 128GB)
Recommended (8-bit): Weight size 9.6GB, generation speed ~18 tok/s.
Mid (6-bit): Weight size 7.3GB, generation speed ~20 tok/s.
Smallest (4-bit): Weight size 5.1GB, generation speed ~22 tok/s.
Key takeaways:
16GB Mac really works: 4-bit 5.1GB, 6-bit 7.3GB — even with half RAM reserved for OS, 16GB fits. This is the core advantage over 4-bit 27B (~17–18GB, won't fit on 16GB).
Speed is "9B-level, not 27B-slow": 18–22 tok/s on M5 Max 128GB feels snappy vs single-digit tok/s of larger models. But your 16GB Mac will be slower due to unified memory bandwidth and thermal limits — netprojects' numbers are not your machine.
Loading requirements: Python 3.11+, mlx-lm>=0.32, and --trust-remote-code (Looped Transformer is custom, not yet in Transformers mainline). For enterprise, treat this third-party code as an external dependency for security review.
⚠️ Version note: The above MLX tests are on the Preview version (netprojects tested Nov 7). The Nov 17 official version is being rebuilt and re-tested; verify the model card date before pulling.
4. The Honest Speed Cost: Looping Is Not Free
Why it's slower: the middle 24 layers run twice → ~1.75× compute per token vs regular 9B. Weight sharing ≠ KV cache sharing — each pass sees different hidden states, so both KV caches or loop states must be kept. You save VRAM, you pay compute time.
Community measurements confirm:
GooGZ AI on RTX 3080 (10GB): SenseNova agent test 30/30, total 226/300 — brainpower is solid; but generation 49 tok/s vs 89 tok/s for regular 9B — halved ; prompt processing 1450 vs 2638 tok/s.
MindStudio on A6000: invoice parsing ~2 tok/s, one structured extraction >3 minutes. Model card shows DSpark/DFlash2 acceleration components still "training"; only MTP available — partly explains slowness.
Preview version occasionally "hallucinated loops" in long reasoning, emitting 32K tokens non-stop.
Accurate description: Loop-9B is "time-for-space" — lets 10GB GPUs tackle 27B-complexity tasks, at the cost of latency. Suitable for latency-insensitive, quality-critical workloads (local doc parsing, agent workflows, offline batch). Not suitable for real-time chat or high-concurrency online serving.
5. Decision Guide: Should Your 16GB Mac Still Chase 27B?
Two paths:
Stick with 27B: 4-bit ~17GB, 16GB Mac can't fit; need 24GB minimum, 32GB comfortable. Gain: stronger general knowledge and depth.
Switch to Loop-9B: 4-bit 5.1GB, 16GB Mac runs easily; specific tasks reach 27B level. Cost: speed halved, general knowledge still below true 27B.
By user profile:
You have a 16GB Mac: Stop forcing 27B — it won't fit and will crawl. Loop-9B MLX 4-bit/6-bit is the best "runs and strong" option today — accept the slowdown.
You have 32GB: Both work. Want best general ability → 27B. Want agent/long-doc/low footprint → Loop-9B.
You prioritize responsiveness: Loop-9B is not for you. Get higher-bandwidth hardware and run standard quantized models; the loop architecture's speed tax is too high.
Enterprise/private deployment: 10GB VRAM floor is attractive; Cambricon day-one support. But --trust-remote-code third-party impl, TBD technical report, unfinished acceleration components — all must enter risk assessment.
6. Data Discipline: What's Solid, What Needs a Question Mark
Hard facts (multi-source verified): Nov 17 official release, 9B params/32 physical layers/56 layer passes, 4-bit 4.9GB, ≤10GB deploy, 512K context, Apache 2.0, Cambricon day-one — consistent across SenseTime site, IT之家, Tencent News, Sohu.
Vendor self-reported (no independent verification): All benchmark scores (NL2Repo, AIDABench, WorkBuddy, etc.) from SenseTime; technical report marked TBD. Treat as directional, not absolute.
Community tests (unofficial, small sample): MLX 18–22 tok/s from netprojects on M5 Max 128GB, Nov 7 Preview (ARC 300, GSM8K 50); author calls few-point differences noise. 49 vs 89 tok/s from GooGZ AI on RTX 3080. Use for magnitude feel, not precise baselines.
Version trap: MLX numbers mostly Preview; official version just released, not yet re-tested; benchmark target shifted from Qwen3.5-27B to Qwen3.8-27B — don't mix.
Precise meaning of "9B runs like 27B": On agent/code/office subset of tasks it reaches 27B threshold, not full equivalence, and speed does not keep up.
(Data as of Nov 18, 2026. Release info/specs from SenseTime official, IT之家, Tencent News, Sohu; official benchmarks are vendor self-reported; MLX speeds from netprojects Nov 7 Preview test on M5 Max 128GB; community comparisons from GooGZ AI / MindStudio. All non-official figures annotated with source and sample limits; rely on your own hardware tests and latest model card.)
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
