Multiverse Computing Launches Quantum-Healed Quasar 1.1

Multiverse Computing has launched Quasar 1.1 438B, a compressed coding model derived from Z.ai’s GLM-5.2. To shrink the model, Multiverse pruned its mixture-of-experts layers from 256 experts down to 148, representing a 42.2% reduction. To repair the reasoning and instruction-following capabilities typically damaged by aggressive pruning, the company used a recovery process it calls healing. This stage retrained the network on synthetic data generated by a hybrid quantum language model, marking the first time physical quantum hardware has been integrated into the CompactifAI training pipeline.

The synthetic training data was generated using a modified Qwen3-30B-A3B model, which had one-sixth of its layers replaced with a quantum neural network. These parameterized circuits ran on the 156-qubit IBM Heron processor inside the IBM Quantum System Two located in Donostia-San Sebastián, alongside a calibrated noise model. While this hybrid setup generated the recovery data, running inference on the resulting Quasar 1.1 model does not require quantum hardware.

Compared to its predecessor, Quasar 1.0, the new model shows notable benchmark improvements, including gains of +6.2 HLE, +4.3 GPQA, +4.6 IFBench, and +6.4 LCR. Furthermore, Quasar 1.1 generated 37.6% fewer output tokens on average across five benchmarks—SciCode, HumanEval, GSM8K, TriviaQA, and BBH—resulting in an average reduction of 1,248.6 tokens per response. This reduction lowers operational costs for token-metered agent systems.

The update also addresses political refusals inherited from GLM-5.2. By applying a method called Refusal Steering, Multiverse lowered political refusals from 63.75% to 41%, a drop of 22.75 percentage points, while maintaining its JailbreakBench safety score at 93% and increasing harmful-prompt refusals by one point. Quasar 1.1 is currently available via the CompactifAI API. To incentivize testing, Multiverse is hosting a bug-bounty challenge that will award 60 million API tokens to each of the top five flaw reporters.