Last updated: September 15, 2026
Short answer: Quasar 1.1 438B is not an AI model running on a quantum computer. Multiverse Computing says that part of the data used to retrain and improve the conventional language model was generated by a hybrid quantum language model using an IBM Quantum processor. The quantum step happened upstream in the training pipeline; developers still access Quasar through a classical API.
That distinction is the entire story. The announcement is interesting because quantum hardware was used to produce data that later shaped a large language model. It is much less accurate to describe Quasar as a quantum LLM, a model that requires a QPU for every answer, or proof that quantum computers have already replaced classical AI infrastructure.
Multiverse Computing announced Quasar 1.1 on September 15, 2026 as a rebuild of its 438-billion-parameter reasoning model. The company says the new release went through another CompactifAI pipeline, including broader healing data, shorter output tuning and a quantum-generated component. The results and the word “first” should be read as company claims until independent testing isolates the effect of the quantum data.
What Is Quasar 1.1 438B?
Quasar 1.1 438B is a reasoning model from Multiverse Computing, a Spain-based company that develops model compression and deployment technology. It is aimed at coding, tool use and enterprise agents. The “438B” label refers to the model’s parameter scale, not to the number of parameters that must be active for every token or to the amount of quantum hardware behind it.
Its lineage matters. Multiverse’s technical write-up says Quasar 438B was built from GLM-5.2, an open-weights model from Z.ai, and then transformed with CompactifAI. The company describes targeted expert pruning, a healing pass and quantization-aware variants. That makes Quasar a European product and deployment offering, but not a European foundation model trained from scratch.
The distinction is useful for readers comparing sovereign AI products. A model can be developed, served and governed by a European company while still inheriting a base checkpoint from elsewhere. In Quasar’s case, the model lineage and the deployment stack are different parts of the product.
What changed in Quasar 1.1?
According to Multiverse, version 1.1 is more than a new label on the same compressed checkpoint. The company says it reran the model through a broader retraining or “healing” set containing reasoning traces, tool-call sequences and general knowledge. It also tuned the model to stop using as many output tokens and changed some refusal behavior inherited from the base model.
| Reported change | What the company says | What it does not establish |
|---|---|---|
| Broader healing data | Reasoning traces, tool-call sequences and general knowledge were added after pruning. | It does not isolate which data source caused each improvement. |
| Quantum-generated data | Part of the healing set came from a hybrid quantum language model. | It does not show that the whole model was trained on a quantum computer. |
| Shorter outputs | Average output fell from 3,322.9 to 2,074.3 tokens across five listed evaluations, a reported 37.6% reduction. | Fewer tokens are not automatically better reasoning or lower end-to-end agent cost. |
| Refusal behavior | Political-topic refusals reportedly fell while harmful-prompt refusal stayed high. | This is a separate steering and post-training change, not evidence of quantum advantage. |
That list is important because it prevents a common attribution error. If Quasar 1.1 scores better or produces shorter answers, the result could come from a combination of data, pruning, healing, quantization, output tuning and refusal steering. A before-and-after headline cannot tell us how much came from the quantum-generated samples.
What Does “Quantum-Generated Training Data” Actually Mean?
In the release, the quantum processor appears in a data-generation stage rather than in the serving path. Multiverse says one sixth of the layers of Qwen3-30B-A3B was replaced with a quantum neural network in a hybrid quantum language model. Circuits were run on IBM Quantum System Two in Donostia-San Sebastian, using a 156-qubit IBM Heron processor and a noise model calibrated from that device.
IBM Heron QPU → hybrid quantum language model → selected synthetic examples → part of the healing dataset → Quasar 1.1 post-training → classical inference through CompactifAI API.
“Generated” does not mean that a quantum computer wrote the entire Quasar corpus. It means the company used a quantum-enhanced model to produce a portion of the examples used in a later training or healing step. The release does not publish the size of that subset, its exact prompts, its filtering rules or an independent audit of its distribution.
That is why Quasar itself does not need a quantum computer to answer your prompt. Once the training process is complete, the resulting weights can be served on conventional AI infrastructure. A customer calling the CompactifAI API sends a normal API request and receives a normal text response. The QPU is part of the upstream experiment, not a dependency of everyday inference.
Is Quasar 1.1 a Quantum AI Model?
It depends on what “quantum AI” is meant to describe. The label covers several very different architectures and workflows.
| Term | Meaning | Does it describe Quasar 1.1? |
|---|---|---|
| Quantum-generated data | Examples, features or samples produced with help from a quantum processor and used later by a classical model. | Yes, according to Multiverse’s release. |
| Quantum-inspired AI | A classical algorithm influenced by quantum or tensor-network ideas but run without a QPU. | Yes for parts of the CompactifAI compression lineage. |
| Hybrid quantum-classical AI | A workflow in which a QPU and classical hardware exchange data during a computation. | Yes for the reported data-generation experiment. |
| Quantum-native LLM | A language model whose main representation and inference run on a quantum computer. | No evidence supports this description. |
The most accurate short description is therefore: Quasar 1.1 is a conventional large language model whose post-training pipeline includes data generated by a hybrid quantum language model. That is narrower than “quantum LLM,” but more technically meaningful.
Why Would Quantum-Generated Data Help an LLM?
The proposed benefit is not magic speed at inference. A quantum component might generate samples with a distribution or structure that is difficult, expensive or simply different to obtain with a classical generator. If those samples cover useful reasoning or tool-use patterns, they could add information to a healing set.
But “different” is not the same as “better.” A useful test would compare at least three conditions: the original Quasar 1.0, a Quasar 1.1-style retraining run with the same classical data but no quantum-generated subset, and the full 1.1 run with the quantum subset. The datasets, training budget, model checkpoint, evaluation prompts and decoding settings would need to remain controlled.
That ablation is not present in the public release. The company reports gains in HLE, GPQA and IFBench, plus shorter outputs, but the same release also describes broader data and several non-quantum interventions. Until the variables are separated, quantum-generated data should be treated as an interesting training technique, not evidence of quantum advantage for LLMs.
How Much of Quasar Is Actually European?
Calling Quasar a European AI model is defensible when the subject is its developer, product and deployment environment. Multiverse Computing is a European company, the new release describes work in San Sebastian and the company presents the model as a European deployment option. That does not mean the underlying foundation model originated in Europe.
Z.ai → GLM-5.2 open-weights base
Multiverse Computing → CompactifAI compression, targeted healing, quantization and Quasar productization
IBM Quantum context → hardware used in the reported hybrid data-generation experiment
The September 3 technical note says expert layers were reduced from 265 to 148 experts per layer. The September 15 release summarizes the reduction as 256 to 148. Because the two official pages differ, this article does not treat either figure as a precise independent measurement. The safer conclusion is that CompactifAI performs targeted expert reduction and then repairs the model for its intended workload.
How Good Is Quasar 438B?
The headline benchmark numbers need labels. Multiverse’s September 2 launch article reported an Artificial Analysis Intelligence Index score of 43 under an earlier comparison and a 15.3-second time to produce 500 tokens including reasoning. The September 15 release reports 1.1-specific gains on selected evaluations, but it does not provide a public independent ablation for the quantum data.
Artificial Analysis currently maintains a Quasar 438B page described as “max, based on GLM-5.2.” At the time checked, it reported a score of 27 on a newer Intelligence Index presentation, 167 output tokens per second, a 1M-token context window, text-only input and proprietary weights. The page does not clearly label the profile as Quasar 1.1, so it should not be used as a direct 1.1 score.
| Metric | Current reading | Interpretation |
|---|---|---|
| Parameters | 438 billion | Scale label reported for the model family. |
| Context | About 1 million tokens | Useful for long documents and agent traces; it does not guarantee better reasoning. |
| Input modality | Text | Artificial Analysis currently lists no image input. |
| API price checked September 15 | $0.60 input / $1.80 output per million tokens | Official CompactifAI listing; prices can change. |
| Independent profile | 27 on the current AA page | Not labelled as 1.1 and not directly comparable with the vendor’s earlier score of 43. |
For production teams, model quality is only one part of the decision. Long context, tool-call reliability, latency, output length, data handling, contract terms and traceability matter as much as a composite leaderboard number. The New Stack’s analysis makes the same practical point: fast token generation does not automatically mean a fast end-to-end agent when tools and repeated model calls dominate the workflow.
Can Developers Actually Use Quasar?
Yes. The CompactifAI API lists Quasar 438B as available through its hosted service, and the official documentation describes an OpenAI-compatible API design. The current listing shows $0.60 per million input tokens and $1.80 per million output tokens, with a first-month-free offer for new users. Verify the live price and model name before budgeting because the page is a changing product surface.
A typical integration places the API behind an application’s model gateway:
WordPress or AI application
↓
OpenAI-compatible gateway or custom adapter
↓
CompactifAI API
↓
Quasar 1.1
For a WordPress site using an AI plugin, the important work is not the brand name in a dropdown. Check whether the plugin allows a custom base URL, model identifier, timeout, retry policy and logging policy. Keep the API key server-side, cap spend, and record which model version answered each request. A model with a one-million-token context window can still create an expensive or slow workflow if the application sends the entire history on every call.
Do not infer from OpenAI compatibility that every OpenAI feature is supported. Tool calls, streaming, reasoning controls, usage fields and structured outputs should be tested against the current CompactifAI documentation before production use.
What the Announcement Does Not Prove
- It does not prove that Quasar 1.1 runs on a quantum computer.
- It does not prove that every response uses IBM Quantum hardware.
- It does not prove that quantum-generated data caused the reported benchmark gains.
- It does not prove a general quantum advantage over classical synthetic-data methods.
- It does not turn Quasar into a European foundation model trained from scratch.
- It does not guarantee that current API prices, context limits or model behavior will remain unchanged.
The announcement is still meaningful. It moves a quantum language-model idea into a commercial model-development pipeline and makes the experiment testable in principle. The next useful evidence is not another superlative. It is a reproducible description of the dataset, the quantum/classical split, the ablation design and the independent evaluation results.
Conclusion: a synthetic-data experiment, not quantum advantage yet
Quasar 1.1 is best understood as a classical 438B reasoning model rebuilt by Multiverse Computing with a hybrid post-training pipeline. The company says that part of the healing data was generated with a quantum language model executed on IBM hardware. That is a real and interesting use of quantum computing, but it is upstream of inference.
So, did we just see the beginning of quantum AI? Perhaps, but the narrower answer is stronger: we saw a new attempt to use quantum-generated data as one ingredient in synthetic-data and model-improvement work. If controlled experiments show that those samples consistently improve capability or reduce classical compute at the same quality level, the idea could become an important branch of AI training. On the evidence public on September 15, 2026, it is a promising experiment, not proof of quantum advantage for LLMs.
FAQ
Is Quasar 1.1 a quantum AI model?
Not in the sense of a language model that runs on a quantum computer. Multiverse says part of its training or healing data was generated by a hybrid quantum language model, while Quasar itself is served as a conventional model through an API.
Does Quasar 1.1 need a quantum computer to answer prompts?
No. The reported quantum step occurs upstream during data generation. Developers use the resulting model through the classical CompactifAI API.
What is quantum-generated training data?
It is data produced with help from a quantum processor and later used in a classical model’s training or post-training process. It does not mean the entire training corpus came from a QPU.
What is Quasar 1.1 based on?
Multiverse’s technical write-up says Quasar 438B is based on GLM-5.2, an open-weights model from Z.ai, and transformed with the company’s CompactifAI compression and tuning pipeline.
Can developers use Quasar 1.1 through an API?
Yes. Multiverse lists Quasar through the CompactifAI API and its documentation describes an OpenAI-compatible API design. Check current model IDs, limits and pricing before deployment.
Does Quasar 1.1 prove a quantum advantage for LLMs?
No. The public release combines quantum-generated data with broader training, healing, output tuning and other changes. A controlled ablation study would be needed to attribute an improvement specifically to the quantum data.
Sources and further reading
- Multiverse Computing, Quasar 1.1 438B announcement.
- Multiverse Computing, Inside Quasar 438B technical write-up.
- Multiverse Computing, Introducing Quasar 438B.
- Multiverse Computing, CompactifAI API and current pricing.
- CompactifAI API documentation.
- Artificial Analysis, Quasar 438B model profile.
- IBM Quantum, hardware and roadmap.
- IBM Quantum Documentation, processor types.
- IBM, European Quantum System Two in Donostia-San Sebastian.
- The New Stack, Quasar 438B and agent latency.
- Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics.
- DecodeTheFuture, AI agent observability.
Research checked September 15, 2026. Vendor claims, benchmark versions, API prices and model behavior can change; verify the linked source before implementation.
