1.76 GB model · Four coherent circuit paths · 2.06× faster decoding Leona-Q 1.0 mini brings quantum-inspired tensor compression and learned quantum-circuit interference into a compact language model based on MiniCPM5-2B. MPO correction preserves structure lost during low-bit quantization. A four-path square circuit adds a learned interference-driven residual during inference. 量子inspiredなMPO圧縮と、4経路の量子干渉回路を組み合わせた小型LLM。圧縮したまま回路を推論に使い、独自の直接MPO計算で生成を高速化しています。 The quantum circuits run as classical statevector simulations; no quantum computer is required. This release uses a custom Linux/CUDA runtime. — Structured compression: 4-bit MLP weights with two-core MPO corrections, 8-bit attention/output weights, and 5-bit embeddings. — Quantum-circuit interference: four paths of eight data qubits, each with three layers of learned RY/RZ gates and CNOT entanglers. Learned path amplitudes and phases feed a square-graph interference readout. — Compact adaptation: 98,784 learned circuit-adapter parameters, attached at two Transformer layers with a 0.1 residual scale. — Direct MPO inference: apply the compressed factors directly during decoding, with fused output computation and an optimized GPU…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: LeonaQuantum
Теги: gguf, minicpm5, quantum-inspired, tensor-network, matrix-product-operator, mixed-precision, custom-runtime, cuda
Лайков: 3 | Загрузок: 144
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
