Can Quantum Math Really Shrink DeepSeek by 95%? A Spanish Startup Has $215M Riding on the Answer.
Multiverse Computing claims 95% AI compression with 2-3% accuracy loss. $215M raised, enterprise API launched March 2026. Validation exists—with caveats.
Published: March 22, 2026 · Reading time: 2 minutes
The Claim Worth Taking Seriously
A startup from San Sebastián, Spain says it can compress DeepSeek, Mistral, and Meta Llama by up to 95% — with only a 2–3% accuracy loss. Not through brute-force pruning. Through quantum-inspired mathematics that rewrites the model's internal structure.
Multiverse Computing launched a self-serve API and consumer app for these compressed models this week. TechCrunch covered it three days ago. This is not vaporware — but it warrants scrutiny.
How the Technology Works
CompactifAI uses quantum-inspired tensor networks — specifically Matrix Product Operators — to decompose weight matrices inside a model's attention and MLP layers.
Think of it this way: instead of demolishing walls to make a building lighter, they redesign the internal framework so the same structure uses far less material. The model isn't lobotomised — it's mathematically recompressed.
Their arXiv paper (2401.14109) on LLaMA-2 7B shows: 93% memory reduction, 70% fewer parameters, 50% faster training, 25% faster inference — 2–3% accuracy loss.
What Independent Research Found
In July 2025, Sopra Steria's sustAIn team published an independent evaluation on arXiv (2507.08836), testing CompactifAI on Llama 3.1 8B. Finding: the compressed model significantly reduced computational resources while maintaining accuracy.
The catch they named explicitly: "We do not have information about the impacts of the compression itself — this would enable one to calculate how long it would take to amortize the compression and when the overall global impact would be positive."
Real validation. Incomplete picture. The energy cost of running the compression itself was not measured.
What They're Not Telling You Upfront
They compress open-source models only. Llama, Mistral, DeepSeek — and OpenAI's gpt-oss-120b, which is OpenAI's open-weight model released under Apache 2.0. Not GPT-4o. Not Claude. Not Gemini. Proprietary frontier models are untouchable.
Open-source already lags proprietary. A 95%-compressed Mistral is a compressed version of a model that already trails GPT-4o and Claude Opus. The performance gap doesn't close — it potentially widens.
The 95% figure needs a footnote. Multiverse's arXiv paper shows 93% on LLaMA-2 7B combined with quantization. "Up to 95%" is the marketing ceiling, not the documented average.
What's Now Live
- CompactifAI API — self-serve enterprise portal with compressed Llama, DeepSeek, Mistral models
- HyperNova 60B 2602 — 50% compression of OpenAI's gpt-oss-120b (61GB → 32GB)
- Axelera AI partnership (March 18) — targeting on-device edge deployment
The edge use case is the sharpest signal here. If compressed models run on-device without cloud dependency, the subscription model every major AI vendor sells starts looking fragile.
Consumer Protection Q&A
Q: Should I switch from ChatGPT or Claude to a CompactifAI model? A: Not yet. CompactifAI compresses open-source models that already trail GPT-4o and Claude in most benchmarks. Best fit today: cost-sensitive enterprise deployments, edge hardware, privacy-first use cases.
Q: Does this make my current AI subscriptions overpriced? A: For individual users, not yet. CompactifAI's current sweet spot is infrastructure operators running LLMs at scale. The consumer disruption comes if compressed edge models become genuinely competitive with frontier models.
Q: Is the 95% independently verified? A: 93% is verified by Multiverse's own arXiv paper. Sopra Steria confirmed maintained accuracy on Llama 3.1 8B — but didn't measure the compression cost itself. More independent testing across more models is needed.
What Happens Next
30 days: Watch API adoption signals beyond the 100+ existing enterprise clients.
90 days: Does anyone compress a model that genuinely matches GPT-4o in a specific domain? That's the proof-of-concept that changes the conversation.
6–12 months: If edge AI gains traction through Axelera, expect a direct response from at least one major AI vendor. Silence would be the tell.
Bottom Line
What to believe: 93% compression on Llama 7B, real enterprise customers (Iberdrola, Bosch, Bank of Canada), genuine independent validation.
What to interrogate: The gap between "up to 95%" and documented results. The missing compression amortization data. And the baseline: compressed open-source ≠ compressed frontier models.
What to watch: Whether on-device AI through the Axelera partnership makes your cloud-dependent AI subscriptions feel overpriced by end of year.
OneHuman will keep tracking this space.
Sources
- TechCrunch, March 19, 2026: "Multiverse Computing pushes its compressed AI models into the mainstream"
- CompactifAI arXiv: arxiv.org/abs/2401.14109
- Sopra Steria independent evaluation arXiv: arxiv.org/abs/2507.08836
- TechCrunch Series B, June 12, 2025
Verified by OneHuman · March 22, 2026
Share This Article
"A Spanish startup claims to shrink DeepSeek and Mistral by 95% using quantum math—and has $215M in VC funding, 100 enterprise clients, and one peer-reviewed paper backing the claim."
"Compress a 7B-parameter AI by 93%, cut memory by 93%, speed up training by 50%, and lose only 2-3% accuracy. If that math holds at scale, your cloud AI bill becomes optional."
"Multiverse compresses Llama, DeepSeek, Mistral—not GPT-4o or Claude. The open-source models they target don't match proprietary performance. Compression doesn't close that gap."
"When AI runs on your phone without cloud dependency, the companies that own the cloud lose leverage. Watch how Big Tech responds to Multiverse's March 2026 API launch."
Related: DeepSeek
- • DeepSeek Raised Prices Up to 1,114%. What You'll Actually Pay Is a Different Number.
- • DeepSeek Warned Prices Are Going Up. It Won't Say How Much or When.
- • While Five Frontier Labs Were Tightening Limits, One Tracked Vendor Cut Prices 75% and Shipped a Capability Upgrade
- • DeepSeek V4 vs. American AI: How a Chinese Lab Is Winning on Price While US Labs Fight Over Benchmarks
- • DeepSeek Upgraded 100M Users Silently — Is Another Wall Street Shock Coming?