Multiverse Computing
Multiverse Computing compresses models with CompactifAI, so the same model costs less to run rather than being replaced by a smaller one.
- Compression is the product: CompactifAI makes an existing model smaller using tensor network methods, rather than swapping it for a different small model.
- Six models on Eden AI: three of their own, Hypernova, Carina and Quasar, plus compressed Qwen and GLM.
- The price range is the argument: Hypernova 60B at $0.04 and Carina 60B at $0.05 per million input tokens sit at the bottom of the catalogue.
- Quasar 438B: a one million token context at $0.60 per million input, which is unusual for a model of that size.
- Test before you assume: compressed routes are not automatically the cheapest for a given third-party model, and the comparison is one model string away.
What is Multiverse Computing?
Multiverse Computing is a Spanish deep-tech company, headquartered in San Sebastián and founded in 2019. Its product is CompactifAI, a model compression platform built on tensor network methods that come out of quantum physics.
That origin matters less than what it does to a bill. Compression takes a model and makes it smaller without retraining it from scratch, which changes the economics of running it.
Multiverse Computing at a glance
| Attribute | Details |
|---|---|
| Provider | Multiverse Computing |
| Product | CompactifAI, a model compression platform |
| Headquarters | San Sebastián, Spain |
| Founded | 2019 |
| Category | Compressed inference for open-weight models |
| Models on Eden AI | 6, three of their own and three compressed third-party models |
| API style | OpenAI-compatible, called through the Eden AI endpoint |
| Regions | EU, US and global depending on the model |
The third option most teams never evaluate
Picking a model usually feels like a choice between two things. Take the flagship and accept the invoice, or take a smaller model and accept that it is worse at the job. Teams spend a lot of time moving along that line.
There is a third option, and it is the one almost nobody tests: the same model, made smaller. Not a different model with fewer parameters, but a compressed version of the one you already chose. If it holds up on your evaluation, the saving is structural rather than a discount.
That is the whole company. Six years of compression work, and a set of models that are the output of applying it.
Multiverse Computing models on Eden AI
Their own models
| Model | Model string | Input / 1M | Output / 1M | Regions |
|---|---|---|---|---|
| Hypernova 60B | compactifai/hypernova-60b | $0.04 | $0.14 | global, US |
| Carina 60B | compactifai/carina-60b | $0.05 | $0.20 | global, EU, US |
| Quasar 438B | compactifai/quasar-438b | $0.60 | $1.80 | global, EU, US |
These are where the compression work shows up most plainly. Hypernova 60B at four cents per million input tokens and Carina 60B at five sit at the bottom of the catalogue's price range, and Quasar 438B pairs a one million token context with a rate well below what a model of that size normally costs.
Quasar is the company's first model of its own rather than a compressed third-party one. Read it as evidence of what the compression work produces when it is turned inward, which is a better argument than any leaderboard position.
Compressed third-party models
| Model | Model string | Input / 1M | Output / 1M | Regions |
|---|---|---|---|---|
| Qwen3.8 27B | compactifai/qwen-3-8-27b | $0.40 | $2.00 | global, EU, US |
| GLM 5.2 | compactifai/glm-5-2 | $1.10 | $3.50 | global, US |
| GLM 5.3 | compactifai/glm-5-3 | $1.10 | $3.50 | global, US |
These are open-weight models you can also get elsewhere, served here in compressed form. Worth checking case by case: compactifai/qwen-3-8-27b@eu is the cheapest EU route to that model in the catalogue on output tokens, while on GLM there are cheaper routes outside the EU. Compare before assuming either way.
When should you choose Multiverse Computing?
Choose it if:
Inference cost is the constraint on shipping a feature and you have not yet tested a compressed model against your own evaluation. That test is cheap and most teams have never run it.
You need a large context at a small rate. Quasar 438B carries a million tokens at $0.60 per million input, which is unusual for that combination.
You want the option pinned to a region. Carina, Quasar and the compressed Qwen all have EU routes.
Consider another provider if:
Your workload is not English or Spanish. Multiverse publish that language coverage for Quasar themselves, so treat it as best-fit guidance rather than a hidden limitation.
You need a published context window before you build. That field is filled for Quasar in the catalogue and empty for the other five.
You are buying purely on the headline rate for a specific third-party model. Compressed routes are not automatically the cheapest, and the catalogue will tell you in one query.
Multiverse Computing pros and cons
| Pros | Cons |
|---|---|
| Compression is a real third option between a flagship and a smaller model, and almost nobody evaluates it | Context windows are not published in the catalogue for five of the six models |
| Their own models are priced far below the market, from $0.04 per million input tokens | Quasar covers English and Spanish, which Multiverse publish themselves, so check language fit first |
| Quasar 438B carries a one million token context at $0.60 per million input | The compressed third-party models are not always the cheapest route to that model in the catalogue |
| EU, US and global routes depending on the model, so you can pin a region where it exists | Young as a model builder, even though the compression platform is six years old |
What can you build?
Features that did not survive their first cost estimate
The most useful case is the feature you already designed and then shelved because the per-call cost did not work at your volume. A compressed model is the one lever that changes that arithmetic without changing the product.
High-volume classification and extraction
At four to five cents per million input tokens, the small compressed models make labelling, routing and field extraction over a large stream cost roughly nothing. That is the work where a frontier model is hardest to justify.
Long-context work on a budget
A million token window at Quasar's rate suits whole-document and whole-codebase analysis, where the alternative is building a chunking and retrieval layer you would rather not maintain.
Why use Multiverse Computing through Eden AI?
Compression only pays if you measure it. The honest test is running your own traffic through the compressed model and the original and comparing the outputs, and that is exactly the test nobody runs when it costs a second account and a second contract.
On Eden AI both sit behind the same key, so the comparison is a change of model string. If the compressed version holds up on your evaluation, the saving is yours. If it does not, you found out in an afternoon.
Frequently asked questions about Multiverse Computing on Eden AI
CompactifAI is Multiverse Computing’s model compression platform. It uses tensor network methods from quantum physics to make a model smaller, so it costs less to run, rather than replacing it with a different smaller model.
Six. Three of Multiverse’s own, Hypernova 60B, Carina 60B and Quasar 438B, and three compressed open-weight models, Qwen3.8 27B, GLM 5.2 and GLM 5.3.
Hypernova 60B is $0.04 per million input tokens and Carina 60B is $0.05, which places them at the bottom of the catalogue’s range. Quasar 438B is $0.60 per million input with a one million token context.
It depends on the model. Carina, Quasar and the compressed Qwen have EU, US and global routes. The two GLM models are global and US. Check the region field on each model before treating any of them as a residency answer.
Yes, it is Spanish, headquartered in San Sebastián and founded in 2019. That is a statement about the company. Where a given model runs is a separate question answered by the region field on that model.
Multiverse publish English and Spanish coverage for Quasar. If your workload is in another language, check fit before committing to it.
They are using Multiverse Computing
Start building with Eden AI
A single interface to integrate the best AI technologies into your products.