Story perspectives

Hybrid Quantum Model Boosts Llama 3.1 8B Accuracy Slightly

5/26/2026

Summary

Story summary

Hybrid quantum-classical model cut Llama 3.1 8B perplexity by 1.4% with only 6,000 extra parameters (0.000075%).
The model fixed answers on Jovian planet rings and a gene-flow query, selecting all ringed planets.
Cayley-parameterized adapters were added to a frozen LLM layer and run on IBM’s 156-qubit Quantum System Two QPU.
Borja Aizpurua of Multiverse Computing led the May 7 arXiv study claiming end-to-end quantum enhancement of a production-scale LLM, noting noise and error-correction challenges.

1 of 1

Story summary

Hybrid quantum-classical model cut Llama 3.1 8B perplexity by 1.4% with only 6,000 extra parameters (0.000075%).
The model fixed answers on Jovian planet rings and a gene-flow query, selecting all ringed planets.
Cayley-parameterized adapters were added to a frozen LLM layer and run on IBM’s 156-qubit Quantum System Two QPU.
Borja Aizpurua of Multiverse Computing led the May 7 arXiv study claiming end-to-end quantum enhancement of a production-scale LLM, noting noise and error-correction challenges.