Quasar 1.1 Uses Quantum-Generated Data to Repair a Pruned 438B Model

Multiverse Computing has released Quasar 1.1 438B, a compressed coding model based on Z.ai’s GLM-5.2. During post-pruning training, Multiverse added synthetic data generated by a hybrid quantum language model. The company says this is the first use of quantum-generated data in its CompactifAI pipeline.

The compression system reduces each mixture-of-experts layer from 256 experts to 148, cutting the expert count by 42.2%. In a mixture-of-experts model, each expert is a specialized parameter block, and a router selects a subset for each token. Removing experts reduces the model’s stored parameters and deployment footprint, although realized cost and latency depend on the serving stack.

Aggressive pruning can damage reasoning, instruction following, and domain knowledge. Multiverse addresses that loss through a recovery stage it calls “healing,” which retrains the pruned network on reasoning traces, tool-call sequences, general knowledge, and the new quantum-generated examples.

Quantum circuits feed the healing set

Multiverse created part of the healing data with a modified Qwen3-30B-A3B model. The researchers replaced one-sixth of its layers with a quantum neural network, whose parameterized circuits acted as trainable transformations inside the language model. Those circuits ran on IBM Quantum System Two in Donostia-San Sebastián, using a 156-qubit IBM Heron processor. Additional runs used a noise model calibrated from the same hardware.

The hybrid model then generated synthetic examples for Quasar’s recovery training. The announcement places quantum hardware in the data-generation stage; it does not indicate that Quasar’s production inference requires a quantum processor.

CompactifAI already used quantum-inspired tensor-network methods to select experts for removal. Quasar 1.1 extends that work by incorporating outputs from physical quantum circuits into the training corpus. Multiverse has not published an ablation that trains otherwise identical models with and without those examples, so the quantum data’s individual contribution to the reported gains remains unmeasured.

Benchmarks rise as outputs shrink

Multiverse reports gains over Quasar 1.0 on reasoning and instruction-following evaluations. It also reports a lower average output length across five coding, reasoning, and question-answering benchmarks.

Measurement	What it tests	Reported change
HLE	Broad, difficult academic questions	+6.2 points
GPQA	Graduate-level scientific reasoning	+4.3 points
IFBench	Instruction following	+4.6 points
Average output length	SciCode, HumanEval, GSM8K, TriviaQA, and BBH	3,322.9 to 2,074.3 tokens, down 37.6%

The output reduction amounts to 1,248.6 fewer tokens per response on average across the measured tasks. Token-metered agent systems could see lower generation costs and shorter loops, especially when one response triggers several subsequent tool calls. The quality gains and token counts come from different benchmark groups, so teams should verify both measures on the same production workload.

A targeted edit lowers political refusals

According to Multiverse, GLM-5.2 carries topic-level refusals and state-aligned framing on some political and historical subjects. Quasar 1.1 applies the method described in Refusal Steering to reduce those refusals while preserving rejection behavior for harmful prompts.

The method uses an LLM judge to score refusal confidence, then fits a ridge-regularized vector representing the refusal-to-compliance direction in the model’s internal activation space. Adjusting the model along that vector targets a specific behavior without requiring a full retraining run.

Prompt set	GLM-5.2	Quasar 1.0	Quasar 1.1
Politically sensitive prompts	71.18% refusal	63.75% refusal	41.00% refusal
Harmful prompts on JailbreakBench	N/A	92.00% refusal	93.00% refusal

Within these test sets, political refusals fell by 22.75 percentage points from Quasar 1.0, while harmful-prompt refusals increased by one point. Broader claims require adversarial evaluation across paraphrases, multilingual prompts, indirect requests, and tool-enabled attacks.

Agent loops are the target workload

Multiverse positions Quasar 1.1 for agentic coding and tool use. Its healing data emphasizes parseable tool calls, multi-step reasoning, executable code, long-context tasks, and concise final answers. The lower output-token count fits workloads where verbose responses increase both latency and the number of tokens carried into later turns.

The company also presents its EU jurisdiction as a deployment consideration. Multiverse is incorporated under EU law and describes the service as compatible with EU AI Act transparency expectations. Compliance, data residency, and processing location still depend on the service’s contracts, hosting regions, retention policies, and subprocessors.

API access comes with open questions

Quasar 1.1 is available through the CompactifAI API. The launch announcement does not provide per-token pricing. Multiverse is also running an evaluation challenge in which the five participants who report the most verified flaws each receive 60 million API tokens. A model listing is available on Artificial Analysis.

A production-grade test plan

1. Coding agents: Measure compilation, test-pass rates, patch acceptance, tool-call validity, retries, and total tokens per completed task.
2. Tool-heavy RAG: Check schema adherence, citation accuracy, retrieval grounding, and recovery from failed tool calls.
3. Sensitive topics: Evaluate factual accuracy alongside refusal rates, using paired political and harmful prompts.
4. Serving performance: Record median and tail latency, output length, concurrency limits, and cost per successful workflow.

Questions the launch leaves open

• Input and output token pricing
• Context-window and maximum-output limits
• Rate limits, batching support, and structured-output guarantees
• Data retention, hosting regions, and training-data policies
• Model-weight availability and versioning commitments
• Benchmark prompts, variance, contamination checks, and quantum-data ablations

Agentic coding systems and tool-heavy retrieval pipelines provide the clearest initial tests because token use, schema validity, and task completion can be measured together. Teams evaluating open discussion of contested political or historical topics should pair refusal testing with factuality and safety checks rather than treating a lower refusal rate as a complete quality measure.