LeonaQ-1

LeonaQ-1 is a quantum-inspired 27B language model based on Qwen3.8-27B. It combines low-bit weights with learned tensor-network corrections and a trained quantum-circuit residual. The complete model payload occupies 10.27 GB.

The release integrates 11 quantum-inspired techniques across training, text generation, model storage, session compression, and auxiliary text analysis. All components run on classical NVIDIA GPUs. The package includes the model weights, trained adapters, tokenizer, and Python runtime.

Run LeonaQ-1 on NVIDIA GPUs provides the CUDA installation commands and inference examples. Run LeonaQ-1 on a Mac describes the portable Metal and CPU text runtime.

October 20, 2026: circuit-preserving GGUF and faster Metal operations

The Q4_K_M GGUF (16.55 GB) retains the trained six-qubit circuit and merges the MPO corrections into the transformer weights. It requires our dedicated llama.cpp extension; stock llama.cpp, Ollama, LM Studio and vLLM are not supported for this custom GGUF. Follow the build and download instructions.

On one A100-SXM4-80GB, the patched engine measured 48.15 tokens/s for 32-token decoding (three repetitions). Full-model Mac throughput has not been measured. The packed PyTorch runtime also includes fused Metal operations for packed weights and the full six-qubit circuit, a persistent interactive mode, streaming output, and separate loading, first-token and decode timing; see MAC.md.

The GGUF preserves the circuit computation, but MPO merging and Q8-to-Q4 requantization change numerical results. Broad quality parity with the original packed model has not been established. Original packed weights and adapter artifacts remain available. Measurements and verification scope.

Release at a glance

Property	Value
Base model	Qwen3.8-27B
Model payload	10,265,779,625 bytes, including the packed base and four auxiliary artifacts
Tensor-network corrections	Matrix Product Operators in 192 feed-forward linear modules
Generation adapter	Trained six-qubit circuit residual after the final transformer layer's MLP
Verified hardware	One NVIDIA A100 80 GB
Runtime	PyTorch with CUDA, Metal, or CPU text generation
License	Apache 2.0

Eleven quantum-inspired components

Model preparation and generation

The packed model combines conventional low-bit quantization with tensor-network corrections and a circuit residual. Targeted training fits these additions while the base model remains frozen.

Component	Implementation
Matrix Product Operators, or MPOs	Learned tensor-network corrections in 192 quantized feed-forward linear modules.
Six-qubit circuit residual	A trained circuit simulation with Ry, Rz, and controlled-Ry gates adds a residual after the final layer's MLP.
Born-Fisher natural-gradient refinement	Circuit training retains parameter updates derived from the geometry of Born probabilities.
Ballistic simulated bifurcation, or bSB	An Ising-inspired solver selects 256 of 1,024 activation examples for circuit-adapter training.
Simulated quantum annealing, or SQA	Annealing selects rounding decisions for the circuit's int4 readout weights.

The MPO corrections and circuit residual contribute during generation. The released circuit artifact incorporates the coreset selection, natural-gradient updates, and SQA rounding.

Model storage and inference state

Three components apply tensor-network and Ising-inspired methods to stored metadata and inference state.

Component	Implementation
Born-MPS metadata coding	A Matrix Product State probability model supports lossless coding of one module's quantization scales and offsets.
Query-weighted Schmidt decomposition	Adaptive tensor factorization compresses recurrent states for session checkpoints.
bSB attention-cache selection	An Ising-inspired solver selects retained key-value entries across all 16 full-attention layers and preserves their original token positions.

Auxiliary text analysis

The runtime includes two trained English sentiment heads and a response-diversity score.

Component	Implementation
Pauli circuit sentiment head	A trained 13-qubit circuit with 52 angles produces 416 Pauli features for sentiment classification.
MPS sentiment head	An eight-site tensor-network residual head uses two-site density matrix renormalization group, or DMRG, updates.
Quantum density-response scoring	A density-matrix-inspired adjustment summarizes text diversity across eight generated response samples.

Measured results

The following measurements describe the released artifacts and their evaluated configurations. Sentiment accuracy uses the project's reused 872-example SST-2 report partition with one training setting.

Measurement	Result	Scope
Pauli circuit sentiment accuracy	94.38%	English SST-2 sentiment classification
MPS sentiment accuracy	93.81%	The same SST-2 partition
SQA circuit artifact size	201,209 bytes	The preceding circuit artifact occupied 374,705 bytes
Retained attention KV tensors	50% reduction	32 MiB to 16 MiB for one 512-token prefix across all 16 full-attention layers, plus 128 KiB of position metadata
Checkpoint restoration	Identical logits over 32 steps	Full-vocabulary logits from the in-memory compressed state and its saved-and-reloaded counterpart

A single A100 80 GB ran text generation, session checkpointing, and auxiliary sentiment classification. Release verification also reproduced the generation tokens and sentiment predictions with code and adapters downloaded from the public repository.

The evaluation records contain the underlying measurements. The release manifest records the software environment and artifact hashes.

Training and provenance

The base model is Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. LeonaQ-1 combines low-bit weight packing with targeted training of the MPO corrections, circuit residual, and sentiment heads.

The runtime verifies the packaged files and binds the adapters and sentiment heads to the same model fingerprint:

9fca585d565a04af3a8093c190e87b9c322da6fe018fc86ca50def01863e025f

License and acknowledgments

LeonaQ-1 is distributed under the Apache License 2.0. The NOTICE records attribution and model modifications.

We acknowledge the Qwen team for the base model and the researchers whose work on tensor networks, quantum-inspired optimization, and efficient language models informs this release.