CompactifAI (Multiverse Computing)

Thesis

CompactifAI is the AI arm of Multiverse Computing, pivoting a European quantum-software business toward model-compression and inference-efficiency infrastructure P1W2. The evidence shows a company simultaneously (1) staffing up ML/LLM/MLOps/platform/DevOps engineering around its Spain HQ P2P3P5P13E7E9E12, (2) standing up an enterprise GTM motion across Europe, the US, and the Gulf P8P16P17P22E8E20, and (3) shipping first-party repos and launch artifacts that anchor a compression-plus-eval research agenda P9P26P27E3W1W5. The through-line is quantum-inspired optimization (Ising/tensor-network) applied to shrinking LLMs, now being commercialized through a CompactifAI API model catalog and a compressed multimodal model (Comet 60B) P26W1W2W5.

Signal desks

Hiring

• Engineering hiring concentrates on ML/LLM and inference infrastructure: Junior Machine Learning Engineer, Machine Learning Engineer (LLM), Senior LLM Engineer, MLOps Engineer, Senior MLOps Engineer (Training & Inference Optimization), Platform Engineer, DevOps (Junior/Mid), Software Engineer, and Senior Data Engineer P2P3P5P1P13E7E9E11E12E19.
• Research leadership is being layered on: Research Director and Research AI Director in San Sebastian, plus Engineering Manager, AI/ML, and intern pipelines (ML and Quantum Software Engineering) P6P7P20P21E6E16E23E29.
• GTM is expanding in parallel: Enterprise Account Executives and Solution Architects across UK, Germany, France, Italy, Nordics, US (California), and Qatar, plus Sales Directors and a VP of Sales AI/LLM P8P16P17P18P19P22P23P24P25E8E13E15E20E21E22E24E25E26E31E32E33E34.
• Partnerships/infrastructure roles signal a compute and cloud motion: AI Infrastructure Partnerships Manager (Inference & Foundry) and Strategic Cloud Partnership Manager (AWS) P4E17E41.
• Location spread: San Sebastian is the core hub, with Barcelona/Madrid/Zaragoza, then distributed EU (London, Paris, Munich, Milan, Sweden/Netherlands), California, Doha, and Toronto for back-office roles P1P2P8P10P11P16P19P22E8E20.

Forks

• No cited evidence in this pack. The orgs repos appear first-party rather than forks of upstream projects (metadata shows Fork: no across the CompactifAI repos) P9P26P27.

Releases

• CompactifAI API official repo exposes slim and base model IDs: DeepSeek R1 0528 Slim, Llama 4 Scout Slim, Llama 3.3 70B Slim, Llama 3.1 8B Slim/Slim Reasoning, Mistral Small 3.1 Slim, and gpt-oss-20b/120b P26.
• Comet 60B launch (Oct 15, 2026): a compressed multimodal model built on NVIDIA Nemotron 3.5 Super VL, halving total parameters from 120B to 60B with 12B active parameters, targeted at agentic workflows W5.
• Full-Chunked-KL-Loss (Python): a CUDA benchmark for chunked knowledge-distillation losses that reduce VRAM usage P9E1.
• LLM-Refusal-Evaluation (Python): an LLM-as-judge refusal benchmark tied to arXiv 2512.16602 and a Hugging Face dataset, vLLM-powered P27E2.
• Block_removal_through_constrained_binary_optimization (Python) and a hands-on workshops repo round out the public footprint E3P28E5.

Talking

• Research narrative: block-removal pruning recast as an Ising-glass problem, beating a baseline by nearly 23 MMLU points, attributed to Roman Orus and team via Hugging Face W1.
• Launch narrative: Comet 60B framed as half the parameters with preserved multimodal capability for agentic builders W5.
• Business/funding narrative: Series C targeting up to ~$570M (EUR 500M) at a reported $1.7B pre-money valuation (announced 27 Jul 2026), plus company-claimed 80-95% model size cuts via tensor networks W2.
• Conference presence: Multiverse Computing attending the AIDAQ AI/Data/Quantum Summit in Berlin (Bitkom) W3.

Shipping

Public artifacts are early and low-volume but coherent. The CompactifAI API catalog is the main product surface, listing slim variants of open models (DeepSeek R1, Llama 4 Scout, Llama 3.3/3.1, Mistral Small 3.1, gpt-oss) alongside their full-size counterparts P26. The flagship compressed model is Comet 60B, built via a CompactifAI pipeline of tensor-network profiling, compression, and a healing stage, and positioned for agentic workflows W5. Research-adjacent repos ship the underlying techniques: chunked KL-loss kernels for VRAM reduction P9E1, block-removal via constrained binary optimization E3, and a refusal-evaluation harness P27E2. Traction on these repos is minimal (0-7 stars across captures), so the shipped signal is technical direction rather than adoption P9P26P27E1E2E3E5.

Research themes

• Model compression and pruning: whole-transformer-block removal framed as Ising/constrained-binary optimization, with the strongest gains reported under aggressive compression W1E3.
• Tensor networks as the compression engine, running on ordinary hardware W2W5.
• Memory-efficient training/loss kernels: chunked forward/full KL-distillation losses to cut VRAM during distillation P9.
• Post-compression recovery: a healing stage to restore knowledge and capability after pruning W5.
• Safety/eval: automated refusal-behavior evaluation using LLM-as-a-judge with confidence scoring and vLLM batched inference P27.
• Quantum lineage: the company is described as the largest quantum software company in the EU, with flagship products CompactifAI and Singularity P1P5.

Hiring & scaling

The pack shows a scale-up in progress, with an employee-count inconsistency worth flagging: some postings say 180+ employees and growing P1, others say 250+ worldwide P4P6P8, suggesting rapid headcount movement between captures. Repeated ML/LLM, MLOps, Platform, and DevOps roles imply a GPU-heavy inference/training platform buildout, reinforced by Senior Data Engineer and IT Systems roles P2P3P5P1P13P11E7E9E12. The GTM hiring wave (AEs and solution architects across five-plus geographies plus VP of Sales and country Sales Directors) points to enterprise commercialization of the compression products, not a research-only lab P8P16P22P25E8E13E20E21E22. Back-office roles (HR, Admin/Finance, FP&A) round out a maturing org P10P12P14P15.

Category implications

• Infrastructure: chunked KL-loss and VRAM-reduction work plus MLOps/inference-optimization hiring indicate a focus on lowering per-token and per-GPU cost, which is a direct play into neocloud/efficiency economics P9E12E7.
• Partnerships: dedicated AI Infrastructure Partnerships (Inference & Foundry) and AWS cloud partnership roles imply a channel and co-sell/cloud-integration strategy rather than pure self-serve P4E17E41.
• Product: the CompactifAI API catalogs slim versions of popular open models (DeepSeek, Llama, Mistral, gpt-oss), positioning it as an inference/compression layer over third-party base models P26.
• Research: compression and pruning are the core moat, with public method artifacts (Ising block-removal, tensor networks, healing) differentiating it from pure hosting providers W1W2W5.
• Hiring: the split between ML/infra engineering and multi-geo sales/solution-architect roles implies parallel buildout of both product and GTM, with San Sebastian as the technical hub and London/Paris/Munich/Milan/Doha/California as commercial front-ends P2P5P16P22E8E20.
• GTM: enterprise targets are explicitly finance, energy, manufacturing, telecom, and industrials across Europe and the Gulf, aligning sales hiring with those verticals P6P16P17P22.

Traction highlights

• Comet 60B received third-party launch coverage framing it as half-parameters/same-power for agentic builders W5.
• A reported Series C targeting up to ~$570M at a $1.7B pre-money valuation (announced 27 Jul 2026) is the largest traction marker, though the sources note reports differ on whether the round has closed W2.
• CB Insights recognition (2023 and 2025) as one of the 100 most promising AI companies is cited in job postings P1.
• GitHub traction is thin: repo stars range from 0 to ~20 in the cited captures, indicating early-stage public artifacts P9P26P27E1E2E3E5.
• Pruning results claimed a ~23 MMLU-point improvement over a baseline under compression W1.