Multiverse Computing has released Quasar 1.1 438B, the first artificial intelligence model built using data generated from a quantum computer. Circuits run on a 156-qubit IBM Heron processor in Donostia-San Sebastian created a portion of the dataset used to rebuild the coding model, marking the first time quantum-generated data has entered the CompactifAI pipeline.
“Quantum computing is not a label on this release, it is part of how the model was built,” the company states; the new version also achieves a 37.6% reduction in average output length across key benchmarks, dropping from 3,322.9 to 2,074.3 tokens.
Quantum Data Integration into CompactifAI Pipeline
The integration of quantum-generated data into artificial intelligence model building reached a new milestone with the release of Quasar 1.1. This marks the first instance of data created using a quantum computer being incorporated into the CompactifAI pipeline, a process previously reliant on entirely classical datasets. Circuits executed on the IBM Quantum System Two processor contributed to a “healing set” used to refine the model’s performance and characteristics, representing a shift from quantum-inspired algorithms to direct quantum data input.
This approach extends beyond simply shrinking the model; Multiverse Computing rebuilt its flagship coding model using the CompactifAI pipeline to broaden its data base, reduce token output, and eliminate unwanted behavioral traits inherited from the original base model. Following pruning, the team retrained the model against an expanded dataset encompassing reasoning traces, tool-call sequences, and general knowledge, resulting in improvements across several benchmarks.
Specifically, Quasar 1.1 438B demonstrated gains in general knowledge and reasoning, with increases of 6.2% on the HLE benchmark and 4.3% on GPQA, and improved instruction-following capabilities, scoring 4.6% higher on IFBench, CompactifAI says. This reduction in output length was achieved, in part, by employing quantum-inspired tensor network methods for expert selection, reducing the number of experts per layer from 256 to 148.
The quantum synthetic data comprised one-sixth of the layers within a Qwen3-30B-A3B model, further demonstrating the extent of quantum integration. This work builds on years of research at the intersection of language models and quantum computing, and represents a move toward using quantum hardware directly in the development of advanced AI systems.
GLM-5.2 to Quasar 1.1: Performance and Refusal Reduction
Quasar 1.1, Multiverse Computing reports, is a rebuilt coding model based on the open-weights GLM-5.2 from Z.ai, and it underwent a full transformation via the CompactifAI pipeline, focusing on reducing unwanted refusals and streamlining output. This involved not only re-healing the model with a broader dataset but also specifically tuning it to produce responses using fewer tokens. The GLM-5.2 model carried inherent restrictions including hard-coded refusals on politically sensitive topics and a tendency towards state-aligned framing.
The team found that these restrictions did not diminish with simple model compression, according to CompactifAI. The reduction in politically-driven refusals is substantial; Quasar 1.1 438B demonstrates a 41.00% refusal rate on sensitive prompts, a marked decrease from the 71.18% exhibited by GLM-5.2 and the 63.75% of the previous Quasar 1.0 iteration.
Importantly, the model maintains a high refusal rate, 93.00%, on harmful prompts as measured by the JailbreakBench benchmark, indicating that safety protocols remain intact. This behavioral shift was achieved through a technique detailed in “Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics,” where an LLM-as-a-judge system isolates and modifies refusal tendencies without impacting overall capability. The resulting model is positioned as a reliable component for deployment, particularly by European organizations operating under EU law and regulatory expectations, including those outlined in the EU AI Act.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
