Hypnos-Q1: Open-Weight 4B LLM With Real Quantum Hardware in the Forward Pass

Hypnos-Q1 is a 4B reasoning model based on Qwen3.5–4B with one architectural modification: a single special token’s embedding is produced by a learned projection of a real quantum measurement from IBM’s ibm_kingston (Heron r2, 156 qubits). Every forward pass that includes this token incorporates live data from a superconducting quantum processor.

Architecture

A new token <|quantum_sig|> is added to the tokenizer at vocab ID 248078. The embedding layer is wrapped with QuantumAwareEmbedding, which intercepts this token and replaces its embedding with the output of quantum_proj: Linear(6, 2560). The 6-dimensional input is a real measurement from ibm_kingston. Total quantum-bonding parameters: ~18K out of 4.66B.

Quantum signature construction

The 6-dimensional input vector encodes OTOC (out-of-time-order correlator) measurements from SYK-like scrambler circuits at depths 1, 2, 3, plus their pairwise absolute differences. This captures both absolute scrambling magnitude and depth-resolved structure.

Each measurement uses:

• 4 qubits
• Dynamical decoupling (XpXm)
• 4-randomization Pauli twirling
• 2048 shots per circuit
• SamplerV2 primitive from Qiskit

Training

64 unique quantum signatures from independent SamplerV2 batches on ibm_kingston were used to train 8000 samples. Full fine-tuning on H100, ~65 minutes, final loss 1.41. The quantum_proj layer was trained jointly with the rest of the model.

Verifiable provenance

All IBM Quantum job IDs are published in quantum_attestation.json:

• Training corpus: d853tcvtjchs73bqs890
• Validation: d85590mgbeec73aooreg

Signature SHA-256 (77097…995409) is also published. The data is independently verifiable through the IBM Quantum dashboard.

Evaluation

| Benchmark | Hypnos-Q1 (4B) | Comparison | | — -| — -| — -| | GPQA Diamond | 79.4 | o1-preview 73.3, gpt-oss-20B 71.5, DeepSeek R1 (671B) 71.5 | | MMLU-Pro | 81.1 | Phi-4 reasoning-plus (14B) 76.0, o1-preview 74.0 | | ParseBench Text Content | 89.8 | Limited public 4B comparisons |

Limitations

This model does not demonstrate quantum computational advantage. The OTOC measurements function as a structured prior in embedding space, not as a source of super-classical compute. Whether real measurements outperform a PRNG baseline with matched statistics on out-of-distribution calibration is an open empirical question. Training the same architecture with classical random 6-vectors as input is the natural ablation.

The base benchmarks are evaluated without <|quantum_sig|> tokens in the prompts, so they measure reasoning capability of the fine-tuned base model. Benchmarks that exercise the quantum injection path are follow-up work.

Deployment

The repository includes:

• Original safetensors with the QuantumAwareEmbedding wrapper for full Resonance-bonded inference (Python + GPU)
• Five GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16) for classical deployment via llama.cpp, Ollama, or LM Studio
• Ollama Modelfile for one-line setup
• quantum_proj.pt separately for wrapper reattachment

The GGUF variants strip the custom embedding wrapper (llama.cpp does not support custom Python modules in the embedding layer), so they behave as a standard fine-tuned 4B reasoning model without live quantum injection.

Reproducibility

All quantum compute was obtained via the IBM Quantum Open Plan (free tier). The pipeline can be reproduced by any researcher with:

• An IBM Quantum account (free)
• A rented H100 for ~1 hour
• A Hugging Face account

Hypnos Q-series

Hypnos-Q1 is the first model in the Hypnos Q-series — Merlin Research’s quantum-classical hybrid LLM line. A companion preprint documenting the broader quantum experiments on the same hardware (45-realization SYK OTOC statistics, operational ER=EPR fidelity-concurrence correlation, discrete time crystal, 40-step quantum Zeno freeze) is in preparation and will be cross-listed on arXiv quant-ph and cs.LG.

Try it

Local inference (classical mode, no quantum injection):

Full repository: huggingface.co/squ11z1/Hypnos-Q1

— -

Released by Merlin Research, Stockholm. All quantum compute obtained via IBM Quantum Open Plan free tier.