1. Headline & intro
Europe’s most interesting AI play may not be chasing the biggest model, but the smallest useful one. Multiverse Computing, a Basque “soonicorn”, is betting that compressed, cheaper-to-run models will matter more to real businesses than frontier-scale experiments. Its new HyperNova 60B release is not just another model drop on Hugging Face; it’s a statement about how Europe plans to compete with U.S. giants.
In this piece, we’ll look at what Multiverse actually released, why compression is strategically important, how this fits into Europe’s sovereignty push, and what it signals for AI infrastructure over the next few years.
2. The news in brief
According to TechCrunch, Spanish startup Multiverse Computing has released a new, free version of its compressed large language model HyperNova 60B on Hugging Face. The model is built by applying the company’s CompactifAI compression technology – inspired by quantum computing – to OpenAI’s gpt-oss-120b.
HyperNova 60B reportedly occupies around 32 GB of memory, about half the size of its base model, while aiming to preserve most of the accuracy and capabilities. The updated release, HyperNova 60B 2602, specifically improves support for tool calling and “agentic” coding scenarios, where inference costs can quickly explode.
Multiverse claims the model outperforms at least some competitors, including Mistral Large 3 from French unicorn Mistral AI, on selected benchmarks. The company already works with enterprise clients like Iberdrola, Bosch and the Bank of Canada, and is reportedly in talks to raise roughly €500 million at a valuation north of €1.5 billion.
3. Why this matters: AI’s real bottleneck isn’t parameters, it budgets
In theory, every company wants frontier-level AI. In practice, most can’t afford to run it at scale. That gap between aspiration and budget is exactly where Multiverse is trying to live.
Compressed models like HyperNova 60B target three chronic pain points:
Inference cost – Cloud GPU time is expensive, and most AI spending in enterprises is already shifting from experimentation to production inference. Cutting memory footprint in half doesn’t just save hardware money; it also enables denser utilization and less overprovisioning.
Latency and UX – Smaller, optimized models mean faster responses using fewer resources. For agentic coding and tool-calling workloads, latency compounds as chains grow longer. Here, compression doesn’t just save euros, it changes what is practically deployable.
Deployment freedom – A 32 GB model suddenly makes on‑premise, sovereign or even edge deployments more realistic for banks, utilities, and governments with strict data policies.
Winners from this move are:
- Cost-sensitive enterprises that need strong capabilities but won’t pay OpenAI-scale prices.
- European public bodies that want “good-enough” AI under their control, not necessarily world‑leading frontier models.
- The open-source and research ecosystem, which gains a free, strong baseline to tinker with.
Who loses? Potentially hyperscale API providers whose margins rely on customers tolerating high inference bills, and any player betting that raw model size alone is a durable moat. Multiverse is implicitly arguing that the real economic moat will be efficiency plus sovereignty, not parameter counts.
4. The bigger picture: From “bigger is better” to “deployed is better”
HyperNova 60B lands at the intersection of several broader trends.
First, there’s the maturation of model compression. Quantization, pruning and knowledge distillation have moved from research curiosities to core infra. Companies like Meta (with smaller Llama variants) and open-source communities have been aggressively shrinking models without destroying quality. Multiverse’s twist is to industrialize this as a product, with a quantum‑inspired marketing edge.
Second, it aligns with the rise of specialized deployment models. OpenAI, Anthropic and others push general‑purpose frontier models via APIs. In parallel, a separate ecosystem is emerging around models that are “good enough” but easy to customize, self‑host or embed. HyperNova sits firmly in this second camp, closer in spirit to Mistral’s smaller models or Llama-based stacks than to GPT‑5.
Third, it mirrors the cloud computing playbook. Early on, everyone obsessed over raw compute. Later, the real money accrued to players who optimized utilization and costs (think AWS Graviton or Databricks). AI is on the same path: the frontier is flashy, but the margin is in optimization, orchestration and compression.
Compared with competitors, Multiverse isn’t trying to be Europe’s answer to OpenAI in terms of scale. Instead, it’s positioning itself as an efficiency layer on top of frontier research, including that of U.S. companies. That is both clever and politically delicate: it sells “sovereign solutions” built, in part, from American base models.
The direction of travel is clear: the industry is shifting from “who has the biggest foundation model?” to “who lets enterprises run powerful AI economically and under their own governance?” On that axis, Multiverse is early—but not alone.
5. The European angle: Sovereignty through compression, not confrontation
European policymakers talk a lot about digital sovereignty, but the reality is that the continent is unlikely to outspend the U.S. or China on raw AI scale. Multiverse’s approach is interesting precisely because it accepts this constraint and turns it into a strategy.
By compressing and packaging models in a way that fits into European infrastructure and regulatory realities, Multiverse is building a bridge between Brussels and Bilbao, not trying to out‑San‑Francisco San Francisco.
Some specific angles for Europe:
EU AI Act compliance – Smaller, more controllable models are easier to document, audit and align with risk‑based regulation. A compressed, self‑hostable model is more compatible with high‑risk sectors like finance and energy.
Public sector procurement – Regions like Aragón and agencies such as Spain’s SETT, both mentioned by TechCrunch, are under pressure to show concrete AI benefits while respecting GDPR and upcoming AI rules. A quasi‑sovereign, cost‑efficient model is politically attractive.
Industrial users – Clients like Iberdrola or Bosch fit the archetype of European industrial champions that want AI tightly integrated with proprietary data and OT systems, often on‑prem. HyperNova’s size makes that plausible without insane hardware builds.
There is a cultural point too: European boards are more risk‑averse and privacy‑sensitive than many U.S. counterparts. An open, inspectable, more easily contained model may simply be an easier sell than a black‑box API from across the Atlantic, even if the latter is slightly more capable.
6. Looking ahead: What to watch next
The next 12–24 months will show whether Multiverse’s strategy scales beyond a few flagship customers.
Things to watch:
Transparent benchmarks – The company claims wins over models like Mistral Large 3. Enterprises will want independent, task‑specific benchmarks (coding, RAG, multilingual reasoning) and TCO comparisons, not marketing slides.
Breadth of compressed offerings – TechCrunch notes that Multiverse plans to open‑source more compressed models in 2026. The key question is whether they become a compression platform for multiple base models (OpenAI today, maybe others tomorrow) or stay tied to one ecosystem.
Regulatory positioning – The EU AI Act’s practical enforcement will begin to bite. If Multiverse can pre‑package compliance (documentation, monitoring hooks, risk tools), it could turn regulation from a burden into a moat.
Funding and independence – A rumoured €500 million round at a unicorn‑plus valuation would put Multiverse in a different league. But it will also come with pressure: investors will expect either rapid revenue growth or a clear platform play, not just bespoke projects.
Edge and device‑side AI – At 32 GB, HyperNova is not a smartphone model, but it points in that direction. If Multiverse continues to shrink strong models, it could eventually power AI in factories, vehicles, and telco infrastructure without round‑trips to U.S. data centres.
The risk is obvious: if frontier models keep improving faster than compression, Multiverse could be stuck forever chasing a moving target. Their bet is that economics, regulation and data‑locality will matter more than a 5–10% gap in abstract benchmark scores.
7. The bottom line
Multiverse Computing’s free HyperNova 60B release is less about hugging Face visibility and more about planting a flag: Europe’s AI edge may come from making powerful models small, cheap and sovereign, not from winning the size race. If they are right, the most important AI companies of this decade won’t necessarily be the ones with the biggest clusters, but the ones that let everyone else use AI profitably. The real question for readers building products today: are you optimizing for raw capability, or for the cost and control profile you’ll still be happy with in three years?
