Multiverse Computing’s Quasar 1.1 used a 156-qubit IBM Heron to generate a small upstream component of synthetic healing data for a compressed classical 438B coding model. It did not train the 438B model end to end on a quantum computer, serve inference on the quantum processor, or demonstrate that quantum-generated data caused the model’s reported improvements.
Key facts
- Quasar 1.1 is a post-trained, compressed model based on Z.ai’s GLM-5.2, according to Multiverse.
- The company says real-device execution took about 90 seconds per token and roughly 25 minutes for a 15-token sentence.
- Most reported quantum-data volume came from GPU statevector simulation: 16,500 samples and about 4.2 million tokens.
- Primary source: Multiverse’s Quasar 1.1 release.
The headline becomes intelligible when the pipeline is separated into steps. Multiverse says it starts from Z.ai’s open-weights GLM-5.2 and applies its CompactifAI process: expert pruning, healing or retraining, output-length tuning, quantization, and refusal-behavior editing. It says the release prunes each layer from 256 experts to 148. That is model compression and post-training on classical hardware, not a new 438B foundation model grown inside a quantum processor.
The IBM device appears only in the synthetic-data stage. Multiverse’s technical note says a hybrid generator began with Qwen3-30B-A3B and replaced eight layers with a multi-head quantum-neural-network block. Thirty-two eight-qubit heads worked on the model’s 2,048-dimensional hidden state through shallow circuits and classical mixers. The company ran circuits on ibm_basquecountry, an IBM Heron r2 device at IBM Quantum System Two in San Sebastián, then used outputs as part of healing data after pruning.
The anchor timing figure reveals why this is not a replacement for GPU training. The technical note says one token took approximately 90 seconds on the real quantum device, so a sentence of around 15 tokens took roughly 25 minutes. The bulk data volume instead came from exact GPU statevector simulation of the same circuits, calibrated with noise characteristics from the hardware. That simulation generated 16,500 samples and about 4.2 million tokens. Multiverse says the real device’s contribution was ‘small in volume.’
An analogy helps. Imagine a large publishing house creates a new style guide using a rare, slow printing press for a handful of examples, then prints the full catalog with conventional machinery. The press contributed examples to the style process; it did not print the catalog. Likewise, the QPU contributed a small data source to a broader healing pipeline. It did not select Quasar’s experts, tune all its parameters, or answer user requests.
The company’s own caveat deserves quotation: the technical note calls the experiment ‘not a shortcut to a better model.’ That is not merely modest wording. Quasar 1.1 also changed broader healing data, verbosity tuning, and refusal steering. Without a controlled comparison that holds those changes fixed and varies only the quantum-generated subset, no one can attribute its reported evaluations to the IBM step. The correct claim is that quantum-generated data entered the pipeline. The claim that it improved the final model remains unverified.
There is a real product behind the carefully bounded claim. CompactifAI’s catalog lists quasar-438b for chat completions, tool calling, and structured output. Artificial Analysis independently lists a functioning served endpoint, identifies its GLM-5.2 base, and labels it proprietary. The model is therefore usable through an API. No public checkpoint was verified, so readers should not expect a downloadable 438B file, disk-size figure, or local VRAM recommendation.
The strongest optimistic case is that quantum hardware might eventually produce unusual training distributions or subroutines worth feeding into classical systems. The strongest skeptical case is that the experiment demonstrates an expensive, slow synthetic-data source whose value was not isolated. Both can be true. This is good early hybrid-systems research precisely because it exposes its role and timing rather than pretending a quantum chip has replaced a data center. The next meaningful result would be a public ablation: same base model, same prune, same healing data volume, same tuning—quantum-derived subset versus a classical control. Until then, treat Quasar 1.1 as a classical compressed model with an intriguing quantum footnote, not quantum-trained AI.
A hybrid pipeline can contain a quantum component without making the final product a quantum model in any operational sense. That nuance is the difference between a useful prototype and a misleading claim of industrial displacement. The public record supports exactly that limited conclusion.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)