Multiverse Computing releases Quasar 1.1 438B, partly trained on quantum-generated data

The story so far, in brief; the legwork, and the full article, stay with the original outlet.

Multiverse Computing has released Quasar 1.1 438B, a rebuilt version of its flagship coding model. It starts from GLM-5.2, an open-weights model from Z.ai, and goes through the company's CompactifAI process, which cuts each layer from 256 experts to 148. The company acknowledges that shrinking a model also makes it worse at its job, so most of the work goes into retraining it to restore what the pruning disturbed.

The new part is where some of that retraining data came from: a hybrid quantum language model, Qwen3-30B-A3B with a sixth of its layers replaced by a quantum neural network, whose circuits ran on a 156-qubit IBM Heron processor in IBM Quantum System Two in Donostia-San Sebastián. The company says this is the first time quantum-generated data has fed its pipeline. Until now, quantum served mainly as inspiration.

The company reports gains of 6.2% on HLE, 4.3% on GPQA and 4.6% on IFBench. Average answers fell from 3,322.9 to 2,074.3 tokens, 37.6% shorter. The political refusals inherited from GLM-5.2 have been removed: refusal on politically sensitive prompts drops to 41.00%, from 71.18% for GLM-5.2 and 63.75% for Quasar 1.0. Refusal of harmful prompts on JailbreakBench edges up to 93.00% from 92.00%, so the model will now say more about history and still nothing about weapons. It is aimed at agent work and at European organisations, served by an EU-incorporated company in line with the EU AI Act.