LeonaQ Orbit 9B
Orbit is a Qwen3.5-9B model with two ring-connected circuit adapters. The backbone is frozen in 4-bit precision, with 197,088 trainable adapter parameters.
Results
Orbit scores 71.02 on the ten-benchmark macro-average. Against a parameter-matched MLP adapter, it records 7 wins, 2 ties, and 1 loss.
Orbit uses 8× fewer adapter parameters than MLP-16, with a higher mean benchmark score.
| Model | Adapter parameters | Mean score | GSM8K |
|---|---|---|---|
| Orbit | 197,088 | 71.02 | 79.98 |
| Qwen3.5-9B, same 4-bit backbone | 0 | 70.96 | 79.91 |
| MLP-2 | 197,088 | 70.54 | 77.03 |
| MLP-8 | 788,352 | 70.54 | 76.72 |
| MLP-16 | 1,576,704 | 70.90 | 77.71 |
| LoRA, rank 6, one site | 98,304 | 70.95 | 77.79 |
| LoRA, rank 6, two sites | 196,608 | 70.87 | 74.98 |
Scores are percentages. The mean gives equal weight to ARC-Challenge, ARC-Easy, BoolQ, HellaSwag, MMLU, OpenBookQA, PIQA, SciQ, WinoGrande, and GSM8K. Evaluation uses each task's full evaluation split, raw choice accuracy, and strict exact match for GSM8K.
All adapters use the same 1,000 training examples, 200 updates, and seed 17. Complete scores.
Run
hf download LeonaQuantum/LeonaQ-Orbit-9B --local-dir orbit
cd orbit
pip install -r requirements.txt
python generate.py --prompt "Explain quantum entanglement in simple terms."
The included loader runs the packed backbone and Orbit adapters in PyTorch on an NVIDIA GPU.
from orbit import load_model
model, tokenizer = load_model(".")
Architecture
Each adapter uses four coherent paths with eight data qubits per path. Three layers of learned RY/RZ rotations alternate with ring-connected CNOT pairs. The readout produces 16 features, which are projected back into the model's hidden representation.
The adapters sit after the feed-forward output projections in transformer blocks 16 and 32. The circuits run through classical statevector simulation.
Training
| Setting | Value |
|---|---|
| Base model | Qwen3.5-9B |
| Backbone precision | Symmetric 4-bit, group size 64 |
| Training examples | 1,000 |
| Updates | 200 |
| Optimizer | AdamW |
| Learning rate | 0.0002 |
| Checkpoint selection | Validation cross-entropy |
License
Apache 2.0. Based on Qwen3.5-9B.