CompactifAI (Multiverse Computing)

Thesis

CompactifAI is the AI-compression and inference arm of Multiverse Computing, a San Sebastián-based deep-tech firm that describes itself as the EU's largest quantum software company and is now squarely commercializing 'hyper-efficient' model slimming P1W2. The evidence shows a two-sided buildout: a research and engineering push on pruning, tensor networks, and memory-efficient kernels W1W2P10, alongside a parallel enterprise go-to-market expansion across Europe, the Gulf, and the US P23P20E20. For a neocloud watcher, the core signal is that CompactifAI is monetizing compression itself — fewer parameters, less VRAM, and 'sovereign' on-prem/cloud/edge deployment — rather than raw foundation-model training, which carries direct inference-economics implications W2W3.

Signal desks

Hiring

• Repeated ML/LLM engineering hiring across seniority — Senior MLE (LLM), Junior MLE, MLE Intern, Senior LLM Engineer, MLE (LLM), and Engineering Manager AI/ML — concentrated in Spain hubs (San Sebastián/Donostia, Barcelona, Zaragoza, Madrid) P1P3P7E9E11E16.
• Infrastructure and platform buildout is signaled by DevOps (Junior/Mid), Platform Engineer, IT Systems Engineer, Senior Data Engineer, MLOps Engineer, and Senior MLOps Engineer (Training & Inference Optimization) P2P4P6P12P14E7E12.
• Cloud/foundry commercialization roles: AI Infrastructure Partnerships Manager (Inference & Foundry) and Strategic Cloud Partnership Manager (AWS) P5E17.
• GTM hiring is broad and geographic: Enterprise AEs (UK, Germany, France, Italy, Nordics, Qatar, US), Sales Directors (UK/DE/FR and Italy), a VP of Sales AI/LLM, and multiple Solution Architects (London, Paris, Munich, Doha) P23P20P24P25P17E8E13E15E20.
• Research leadership roles (Research Director, Research AI Director) in San Sebastián point to a formalized research org P21P22.
• G&A scaling (FP&A, controller, HR, and an Administration & Operations Manager in Toronto) runs alongside headcount messaging that shifts between '180+' and '250+ employees' across job posts P15P16P11P13P1P5.

Forks

• No cited upstream fork activity in this pack: every observed GitHub repo under the CompactifAI org is self-owned and marked 'Fork: no' P10P27P28. The public code surface is first-party tooling rather than inspected or adapted third-party frameworks.

Releases

• The CompactifAI official repository publishes 'slim' variants of open models (DeepSeek R1 0528, Llama 4 Scout, Llama 3.3 70B, Llama 3.1 8B, Mistral Small 3.1, GPT-OSS 20B/120B) served through a Completions/Chat Completions API P27.
• Comet 60B launched Oct 15 2026: a compressed multimodal model built on NVIDIA Nemotron 3.5 Super VL, halving parameters 120B→60B for agentic workflows W4.
• A Hugging Face technical account (Sept 21) reports a pruning method beating a baseline by nearly 23 MMLU points W1.
• Full-Chunked-KL-Loss: a CUDA benchmark/kernel for chunked knowledge-distillation losses that 'massively reduces VRAM usage' P10E1.
• LLM-Refusal-Evaluation: a vLLM-powered library (arXiv 2512.16602) for LLM-as-a-judge refusal detection P28.

Talking

• RuntimeWire frames the pruning work as recasting model shrinking as a statistical-physics problem (Ising optimization), attributed to the physicist co-founder W1.
• QNT coverage frames CompactifAI as tensor-network compression running on ordinary computers, with company-claimed 80–95% size cuts (explicitly labeled a company claim), and a 27 July 2026 Series C targeting ~$570M at a reported $1.7B pre-money valuation W2.
• Marcel Pohle's LinkedIn positions CompactifAI as a 'European sovereign AI lab', claims up to 80% compression 'with zero loss in accuracy', and touts on-prem/cloud/edge deployment and Comet 60B W3.
• StayNobleHub details the Comet 60B pipeline (tensor-network profiling → compression → healing) for multimodal/agentic use W4.

Shipping

Concrete public artifacts: the CompactifAI API model catalog P27, the Comet 60B model launch W4, the Full-Chunked-KL-Loss kernel P10, the LLM-Refusal-Evaluation library with a linked arXiv paper P28, and a block-removal repo using constrained binary optimization E3. The Hugging Face pruning write-up is a technical publication rather than a model card, but is the cited evidence for the nearly-23-MMLU-point result W1. Repo traction is thin — GitHub metadata snapshots show 0–6 stars on the first-party repos P10P27P28 — so the shipping signal is the code/API surface itself, not developer pull.

Research themes

• Pruning/compression recast as statistical physics (Ising optimization) W1.
• Tensor-network-based model compression on classical hardware W2W4.
• Block-level removal via constrained binary optimization E3.
• Memory-efficient knowledge distillation (chunked KL losses to cut VRAM) P10.
• Post-compression 'healing' with targeted data (single-turn math/science/reasoning and multi-turn agentic trajectories) W4.
• Safety/refusal evaluation via LLM-as-a-judge P28.

Hiring & scaling

Spain is the engineering core (San Sebastián/Donostia, Barcelona, Zaragoza, Madrid, Guadalajara) P1P2P3P4P6P12P14, with a research nucleus in San Sebastián (Research Director/AI Director plus intern roles) P21P22P7P8. GTM is distributed across London, Paris, Munich, Milan, the Nordics, Doha, and California P23P20P24P25P9P17E8E20. Two infrastructure/commercialization threads stand out: (1) a hiring ramp in MLOps/DevOps/Platform/IT/Data Engineering implying inference-platform and deployment investment E7E12P2P4P6P12P14, and (2) explicit AWS and inference/foundry partnership roles signaling a channel/cloud strategy E17P5. Back-office hiring (FP&A, controller, HR, Toronto operations) suggests a scaling, funding-adjacent company P15P16P11P13. Headcount claims are inconsistent across postings ('180+' vs '250+'), a minor but notable messaging gap P1P5.

Category implications

• Infrastructure: CompactifAI's value proposition is compression — VRAM reduction P10, tensor networks on 'ordinary computers' W2, and 80–95% size cuts (company claim) W2 — which directly targets inference economics for neocloud operators, though no vendor or revenue relationship is cited.
• Product: the API model catalog of 'slim' open-model variants P27 plus a compressed multimodal model (Comet 60B, 120B→60B) W4 positions the company as a model-efficiency supplier rather than a foundation-model trainer.
• Research: the pipeline is inspectable across profiling → compression → healing W4, with separate repos for block removal E3 and knowledge-distillation memory efficiency P10.
• Hiring: the dense MLOps/inference-optimization and cloud-partnership roles E12E17P5 imply a shift from research toward operated, partnered deployment.
• GTM: sovereign-AI framing with on-prem/cloud/edge deployment W3 plus a European sales footprint P23P24P25 suggest an enterprise data-sovereignty motion; the Qatar and US roles P17E20 extend beyond Europe.

Traction highlights

• Funding: Series C targeting up to ~$570M at a reported $1.7B pre-money (announced 27 July 2026) W2.
• Recognition: CB Insights AI 100 in 2023 and 2025 P1.
• Launch: Comet 60B (Oct 15 2026) multimodal/agentic compression release W4.
• Benchmark: nearly-23-MMLU-point improvement over baseline in the pruning write-up W1.
• GitHub pull is weak: self-owned repos at 0–6 stars in page snapshots P10P27P28, with the Full-Chunked-KL-Loss event log showing 20 E1 — low developer traction relative to the funding narrative.