Leona-Q 1.5 Mini

Small footprint. Quantum circuits in the loop.

Leona-Q 1.5 Mini brings trainable quantum circuits into a compact language model's forward pass. Built on MiniCPM5-2B's official Q4_K_M weights, it adds a learned butterfly circuit that runs during both prompt processing and generation.

Four coherent paths. Eight data qubits per path. Three layers of learned rotations and butterfly entanglers. A compact circuit adapter—just 395 KB—adds quantum-state features to the model's internal representations.

This is a hybrid quantum-classical architecture executed through classical statevector simulation. It runs on conventional hardware; no quantum computer is required.

Benchmark results

All 13 evaluated benchmarks are shown below, using the same released butterfly checkpoint and the official Q4 baseline.

Exact scores and sample counts

Scores are percentages with correct/total counts. ARC-Challenge, ARC-Easy, BoolQ, PIQA, HellaSwag, MMLU and WinoGrande use fixed exploratory subsets, not full leaderboard evaluations. MMLU uses five-shot prompts and five questions per subject. OpenBookQA and SciQ use length-normalized accuracy (acc_norm); COPA and RTE use raw accuracy.

AIME and LiveCodeBench use one generated answer per problem with matched settings and no additional output-token cap beyond the native 131,072-token total context window. LiveCodeBench uses the 175 new v6 problems, not the cumulative v6 set. Other tasks use multiple-choice likelihood scoring.

The chart compares the trained quantum-circuit model with the official Q4 baseline, measured in our Linux/CUDA evaluation runtime. See evaluation details for the protocol and statistical analysis, and runtime validation for the portable implementation.

Inside the model

Open the editable draw.io diagram · View full-size SVG

Q4 language model → learned quantum circuit → residual update → next transformer layers

• Butterfly connectivity: CNOT connections span qubit distances 1, 2 and 4 across successive stages.
• Learned quantum gates: RY/RZ rotations transform input-dependent quantum states.
• Coherent readout: data-qubit and cross-path correlations produce features for the language model.
• Inference-time execution: the circuit stays active when generating responses.
• Compact deployment: approximately 1.56 GB of base weights plus a 395 KB circuit adapter. Runtime memory depends on context length.

The base model stays frozen. The adapter was trained on 1,000 examples: 800 general instruction examples and 200 DeepSeek-R1-derived reasoning examples, using answer-token supervision.

Run locally

This repository contains the circuit weights and a source runtime extension. The official base GGUF is downloaded separately and verified automatically.

Use the included Leona runtime. Stock llama.cpp, LM Studio and Transformers do not execute this adapter as-is. Linux/CUDA checkpoint evaluation is complete; full-model macOS and Windows validation is pending.

Install Python 3.12+, CMake, a C++17 compiler and the Hugging Face CLI, then:

hf download LeonaQuantum/leonaq-1.5-mini --local-dir leonaq-1.5-mini
cd leonaq-1.5-mini
python fetch.py
python build.py --backend metal
python launch.py --bin-dir build-metal/bin --model MiniCPM5-2B-Q4_K_M.gguf --server

Open http://127.0.0.1:8080 to chat.

For Windows with an NVIDIA GPU, use --backend cuda and build-cuda/bin; use a Visual Studio developer terminal with the CUDA toolkit installed. CPU and Vulkan build options are also available. These are build targets, not verified platform compatibility claims.

The launcher defaults to a 4,096-token context. Set your own context size according to available memory. Add --circuit off to compare with the official Q4 baseline.

Build and usage guide · Evaluation protocol · Training and attribution

---

LeonaQuantum · Exploring language models with quantum circuits.