Multiverse Computing announced a €189 million Series B—described as approximately $215 million—on June 12, 2025. Bullhound Capital led the round, which the company says brought total funding to about $250 million. The money is intended to scale CompactifAI, a quantum-inspired model-compression platform that can make some open-weight AI models substantially smaller.
Multiverse advertises reductions of up to 95% in model size. Its materials also describe lower memory requirements, faster inference and lower costs, but the results vary by model and benchmark. The strongest savings and performance figures remain company claims rather than independently audited results.
What Multiverse Computing raised
The funding was announced on June 12, 2025. Bullhound Capital led the €189 million Series B, with participation from HP Tech Ventures, SETT, Forgepoint Capital International, CDP Venture Capital, Santander Climate VC, Quantonation, Toshiba and Capital Riesgo de Euskadi–Grupo SPRI. Multiverse said the round lifted its cumulative funding to approximately $250 million.
The company said it would use the capital to expand CompactifAI and its commercial distribution. No valuation was disclosed in the cited announcement. Multiverse’s funding announcement is the source for the round size, investor list and intended use of proceeds.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
What CompactifAI does
CompactifAI is a model-compression system. Multiverse says it applies tensor-network techniques to represent large neural-network models more compactly, producing smaller versions that can run on conventional CPUs, GPUs and potentially edge hardware.
Quantum-inspired does not mean quantum hardware
The “quantum-inspired” description refers to mathematical ideas associated with quantum information. The reported deployments do not require a quantum computer. A customer can use the resulting models in cloud, private-cloud, on-premises or edge environments.
Compression is not the same as quantization
Quantization lowers numerical precision, such as moving from higher-precision weights to 8-bit or 4-bit representations. Compression can change the model’s representation or structure more broadly. A fair evaluation should compare CompactifAI with quantized versions of the same model, not assume the methods are interchangeable.
Rank #2
What “95% smaller” means—and does not mean
The headline figure refers to a reduction in the model representation or size for particular models. It does not automatically mean 95% fewer capabilities, 95% lower latency, 95% lower hardware spending or 95% lower total application cost.
Free tools Windows power users keep installed
One-click scans. No signup required.
The current AWS Marketplace listing describes Slim variants as offering up to 95% size reduction, up to 2× faster inference and up to 50% lower inference costs, with an average precision drop of about 3%. Multiverse’s 2025 announcement described approximately 2%–3% accuracy loss. TechCrunch reported earlier company claims of 4×–12× faster inference and 50%–80% lower inference costs. Those differences may reflect distinct models, software versions or test conditions; they are not a universal performance guarantee.
| Metric | Public figure | How to read it |
|---|---|---|
| Model size | Up to 95% smaller | Maximum vendor claim; not every model or workload |
| Precision or accuracy | About 2%–3% loss; AWS says average 3% precision drop | Average figures can hide larger failures on specific tasks |
| Inference speed | Up to 2× on the current AWS listing; 4×–12× in earlier company claims reported by TechCrunch | Benchmark-, hardware- and concurrency-dependent |
| Inference cost | Up to 50% on AWS; 50%–80% in earlier reported claims | Vendor-reported and model-specific |
Buyers should ask whether the comparison uses the same precision, tokenizer, context length, hardware, batch size and serving engine. They should also test domain, language, long-context and reasoning performance rather than relying on an average benchmark score.
Which models are available?
Models named at the 2025 announcement
The initial coverage identified compressed versions of Llama 4 Scout, Llama 3.3 70B, Llama 3.1 8B and Mistral Small 3.1. Multiverse said it planned to add DeepSeek R1 and other open-source and reasoning models.
Current catalog and access
As of August 18, 2026, the CompactifAI API catalog includes original and Slim models from Mistral, Qwen, NVIDIA, Z.ai and Multiverse, as well as open-weight OpenAI GPT-OSS models. Open-weight GPT-OSS models are not the same as access to OpenAI’s proprietary hosted API models.
The catalog is dynamic. Examples observed on that date included Mistral Small 3.1 at $0.11 per million input tokens and $0.17 per million output tokens, Mistral Small 3.1 Slim at $0.05 and $0.08, HyperNova 60B at $0.04 and $0.14, GPT-OSS 120B at $0.05 and $0.23, and Whisper Large V3 Turbo Slim at $0.000134 per minute. Prices can change.
How smaller models could reduce spending
- Memory: Smaller weights can fit in less GPU or system memory, potentially allowing cheaper hardware or more concurrent requests.
- Throughput and latency: Faster generation can reduce the machines needed for a target traffic level and improve response times.
- Energy: Less computation may lower electricity use in high-volume serving.
- Bandwidth and storage: Smaller files are easier to distribute to remote sites and devices.
- Edge operation: Local inference can reduce recurring cloud calls, network delay and some data-transfer exposure.
These benefits apply to inference, not the entire AI-product budget. Storage, networking, monitoring, data processing, support, engineering, safety checks and human review remain. A model with a lower token price may cost more overall if quality losses cause retries, longer prompts or additional verification.
Deployment options
Cloud API
CompactifAI is available as a usage-based API through its own service and AWS Marketplace. The AWS product is designed as a serverless access layer; additional AWS infrastructure charges may apply. This is the simplest route for teams that do not want to operate model servers.
Private cloud and on-premises
Multiverse has described on-premises licensing and private endpoints. These options can help with data governance and network control, but the buyer carries hardware, operations, upgrades and validation responsibilities.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Best Value
Edge devices
The company has discussed PCs, smartphones, cars, drones and Raspberry Pi-class devices as potential targets. Actual suitability depends on the exact device, runtime, memory, thermal limits and model quality; a cloud benchmark does not establish edge performance.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the public evidence establishes
Established facts
- The Series B, date and investors are documented in Multiverse’s announcement.
- CompactifAI is a commercially offered compression product with an AWS Marketplace presence.
- Public pages list model-specific prices and Slim variants.
- The compression, quality, speed and savings figures are published company claims.
Still unproven by the cited material
- Independent replication of the maximum size, speed or cost claims.
- Uniform 95% reductions across all supported models.
- Production customers achieving the advertised savings.
- Equal performance on every safety, multilingual, reasoning or specialized-domain task.
- A complete public comparison with quantization, pruning, distillation, vLLM or TensorRT-LLM.
The reviewed AWS listing showed zero customer reviews for the referenced product listing. That does not disprove the technology, but it limits public buyer validation.
How it compares with other efficiency strategies
| Approach | Strength | Trade-off |
|---|---|---|
| Quantization | Widely supported and often straightforward to deploy | Can reduce numerical quality; gains depend on hardware and runtime |
| Distillation | Can tailor a smaller model to a task | Requires training data and may lose capabilities absent from that data |
| Pruning and sparsity | Removes parameters or exploits zero structure | Sparsity is not automatically faster without suitable hardware and software |
| Inference engines | vLLM, TensorRT-LLM and ONNX Runtime can improve serving without changing weights | Require deployment expertise and may be complementary rather than substitutes |
| Smaller native model | Often the simplest path to low latency and cost | May lack the broader capabilities of a compressed larger model |
Who should consider it?
Potentially good fits
- High-volume inference where small per-request savings compound.
- GPU-memory-constrained or latency-sensitive services.
- Private or edge deployments that can tolerate measured quality changes.
- AWS customers wanting consolidated procurement and billing.
- Teams using supported open-weight models.
Likely poor fits
- Applications requiring exact parity with a reference model.
- Medical, legal, financial or scientific systems without domain-specific validation.
- Proprietary hosted models that cannot be exported or compressed.
- Low-volume workloads where testing and integration cost exceed infrastructure savings.
- Strictly regulated or safety-critical systems unable to tolerate small regressions.
A practical buyer test
- Choose representative production prompts, including failures, long context, multilingual and reasoning cases.
- Run the current model, the relevant CompactifAI Slim model, a similarly quantized version, a smaller native model and the existing serving engine.
- Measure task success, refusal and safety behavior, time to first token, tokens per second, concurrent throughput, memory use and cold-start latency.
- Calculate cost per successful completed task, including retries, human review, hardware, storage, networking and engineering time.
- Review the underlying model license, data handling, private-endpoint terms and regional availability before deployment.
Do not treat a token price or a maximum compression percentage as a total-cost estimate. The AWS Marketplace page also notes that AWS infrastructure charges may be additional.
Bottom line
The financing signals strong investor interest in making AI inference more efficient, and CompactifAI is a real commercial product rather than only a research proposal. Smaller representations could matter most where GPU memory, latency, energy or edge connectivity are binding constraints. The unresolved question is breadth: whether the company’s headline savings and roughly 2%–3% quality trade-off hold for a buyer’s exact model, hardware and production tasks.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
