Multiverse Computing

Yesterday, our Co-founder and Chief Scientific Officer, Román Orús, participated in the Spain–France Trade and Investment Forum in Madrid, organised by ICEX alongside CEOE, the Spanish Chamber of Commerce, Business France and MEDEF. Held on the occasion of French President Emmanuel Macron’s visit to Spain, the forum brought together business and institutional leaders to strengthen trade, investment and partnerships between the two countries, with European competitiveness and technological sovereignty high on the agenda. 🎙️ Román joined the roundtable “Spain and France facing the challenge of European strategic autonomy: frontier technologies”, alongside: 🔹 Christel Heydemann, CEO of Orange 🔹 Emmanuel Le Roux, CEO of Bull 🔹 Ana María Molina, CEO of Hisdesat Servicios Estratégicos 🔹 Jose SAINZ Armada, Chairman of Iberdrola España The session was moderated by Elisa Carbonell, CEO of ICEX, and focused on the role of frontier technologies in advancing Europe’s strategic autonomy. At Multiverse Computing, we believe that building Europe’s technological capabilities requires both scientific innovation and strong cross-border partnerships. We are pleased to contribute to this conversation and to the development of efficient AI that gives organisations greater control over their technology. #MultiverseComputing #ArtificialIntelligence #EuropeanInnovation #SpainFrance #StrategicAutonomy

🤝 Bringing specialised AI to unattended payments and regulated financial services. Multiverse Computing and Muxunav have formed a long-term partnership to develop a new family of AI models, starting with deployment in Muxunav’s unattended payment devices. The collaboration focuses on three areas: 🎙️ Voice assistance: helping customers understand products, complete purchases and resolve issues, directly from the payment device. 🔎 Identity verification (KYC): combining document reading, facial biometrics and screening to support customer onboarding, with human review where appropriate. 🛡️ Fraud prevention: analysing transaction, device and behavioural signals to identify risks in real time and flag or block suspicious payments. Designed for environments with limited computing resources and variable connectivity, these models will bring AI into everyday services, from vending machines and laundries to EV charging stations. Our shared goal: make customer assistance more accessible while keeping speed, data control and operational requirements at the centre of deployment. Starting with unattended commerce, the partnership will extend these capabilities to regulated financial services. 👉 Learn more: https://lnkd.in/erT6CsyM #MultiverseComputing #Muxunav #EfficientAI #EdgeAI #Fintech

⚽ Bringing advanced AI to the heart of football. We are joining forces with Sports Data Campus to help clubs, leagues and federations turn sports data into actionable insights. This strategic alliance combines our expertise in efficient, specialised AI with Sports Data Campus’s educational programmes and international network, connecting technology with the professionals who will put it into practice. Together, we will explore opportunities across: 📊 Performance analysis and scouting. 🏃 Talent development, workload and injury risk management. 🤝 Fan engagement. 🎓 Education and applied projects that equip sports professionals to use AI effectively. Building on our Player Injury Risk Monitor and our work with organisations such as Real Sociedad, this collaboration opens new paths for AI adoption across the global football ecosystem, with a strong focus on the Americas. 🔒 Our approach enables sports organisations to deploy advanced AI within their own infrastructure while retaining control over their data. We look forward to working with Sports Data Campus to bring technology, knowledge and practical applications closer together. 👉 Learn more: https://lnkd.in/ek89KZzA #MultiverseComputing #SportsDataCampus #AI #SportsAnalytics #Football

Most factuality checks for LLM agents ask whether a claim is supported somewhere in the evidence. When an agent pulls from search tools, databases, and records through MCP, that is not enough. A refund window can be real, live in the policy document, and still make the answer wrong, because the answer says "according to your account record". We call this failure cross-source conflation. In our latest paper we introduce ProvenanceGuard, a verification layer that checks each claim against the source the agent actually pulled it from, and separately whether that source is the one the answer names. On 281 real medical-agent traces, human experts flagged 139 claims that should not pass. ProvenanceGuard caught 138, scoring highest of the five support checkers we compared. The approach is already out of the lab. An optional grounding-verification stage using ProvenanceGuard's source-aware verification was merged into NVIDIA's NVFlow, where it checks a finance agent's completed answers against the SEC excerpts it retrieved. The paper was also presented as a poster at the Agentic AI Summit 2026 at UC Berkeley. Paper, full writeup, and the NVFlow integration linked in the comments. #LLM #AIAgents #MCP #Factuality Ander Alvarez Sanz Santhiya Rajan Alessandro Genuardi Oquendo Oliver Wirjadi Samuel Mugel, PhD Román Orús Multiverse Computing

Compressing a large language model by deleting whole transformer blocks is one of the cheapest ways to make it faster, since a shorter model runs faster and is easy to combine with quantization or other compression. The hard part is choosing which blocks to remove. Blocks interact: removing block 20 hurts more or less depending on whether you also removed block 19 or 24, so it's a combinatorial problem, not a ranking problem. In one of our recent papers, we map this to a problem straight out of physics. Each block gets a binary variable, keep or remove, and a second-order expansion of the model's loss gives us a Hessian whose off-diagonal terms are exactly the interactions between blocks. This turns block selection into finding the lowest-energy state of an Ising glass, and that energy correlates strongly with how well the pruned model actually performs. Once the couplings are computed, checking a candidate is one cheap calculation, no need to run the model, and the hardest cases can be handed to the same quantum-inspired solvers used across our work at Multiverse. The gains show up most at deep compression. At 50% compression of Llama-3.3-70B-Instruct, we get close to 23 points higher MMLU than the best competing block-removal method, and the approach generalizes cleanly to heterogeneous architectures like NVIDIA's Nemotron MoE model. Paper and code linked in the comments. #LLM #ModelCompression #AIResearch #QuantumInspired David Jansen Roman Rausch Román Orús Ali Hashemi

🗒️ 🎙️Enterprise speech-to-text, powered entirely by CPUs. Together with Hewlett Packard Enterprise and Intel, we have demonstrated how we can bring high-volume transcription to standard enterprise servers, with no GPUs required. Our optimized version of Whisper Large V3 Turbo runs on a single HPE ProLiant DL380 server with two Intel Xeon 6 processors, delivering: ⚡ Approximately 2× the throughput of the original model on the same hardware. 🎙️ Capacity equivalent to around 17,000 hours of audio transcribed per day. 📉 50% fewer parameters and half the memory footprint. 🎯 Word error rate below 3% in our benchmark. 🔒 All audio stays within the customer’s own infrastructure. For organisations managing large volumes of recorded conversations, this opens up opportunities to transcribe more customer calls, build searchable archives and expand quality monitoring while retaining control over their data. This is the first result of our ongoing collaboration with HPE and Intel, demonstrating how efficient AI models can unlock advanced workloads on infrastructure enterprises already operate. 👉 Explore the results and learn more: https://lnkd.in/eW3BGw6b #MultiverseComputing #CompactifAI #HPE #Intel #EfficientAI #SpeechToText

🤝 AI Partner Summit: Shaping the future of AI together in Donostia-San Sebastián. Yesterday, we hosted our inaugural AI Partner Summit at Palacio Miramar, bringing together senior executives from our strategic European partners to share perspectives, align on collaborative workstreams and shape our joint AI roadmap for the years ahead. The programme explored key priorities for our collaboration: 🚀 Our vision for AI at scale and the role of Multiverse Computing. 🇪🇺 Progress on IPCEI and opportunities for European collaboration. 💡 Partners’ perspectives on enterprise AI and the future of Europe's sovereignty 🗺️ Our shared roadmap, priorities and next steps. From partner sessions and roundtable discussions to conversations over coffee, the summit provided a valuable opportunity to strengthen relationships and explore how our complementary expertise can support Europe’s AI ambitions. Thank you to all our partners and institutional representatives for joining us and contributing their insights. We look forward to turning these conversations into joint action. #MultiverseComputing #AIPartnerSummit #ArtificialIntelligence #EuropeanAI #Collaboration

🚀 Introducing Luminary, our AI wind tunnel for enterprise model selection. 2.9M LLMs in the market. Which one performs best for your business? The answer depends on how it handles your workflows, customer conversations and policies. Luminary simulates hundreds of complete, multi-turn customer conversations, including backend systems and tool calls, to help you select the right model for your chatbot or agentic system. 📊 Compare models on cost per completed task, success rate, latency and policy compliance. 🔍 Identify lost context, tool misuse and policy drift before deployment. ⚖️ Evaluate Multiverse Computing models alongside third-party models from any OpenAI-compatible inference provider. 📋 Reproduce each test, with an exportable audit trail to support your decisions. As new models become available, rerun evaluations to assess upgrades and find opportunities to reduce costs. 👉 Check it out: https://lnkd.in/eth6RRE8 #MultiverseComputing #Luminary #EnterpriseAI #AgenticAI

🚀 Introducing Luminary, our AI wind tunnel for enterprise model selection. 2.9M LLMs in the market. Which one performs best for your business? The answer depends on how it handles your workflows, customer conversations and policies. Luminary simulates hundreds of complete, multi-turn customer conversations, including backend systems and tool calls, to help you select the right model for your chatbot or agentic system. 📊 Compare models on cost per completed task, success rate, latency and policy compliance. 🔍 Identify lost context, tool misuse and policy drift before deployment. ⚖️ Evaluate Multiverse Computing models alongside third-party models from any OpenAI-compatible inference provider. 📋 Reproduce each test, with an exportable audit trail to support your decisions. As new models become available, rerun evaluations to assess upgrades and find opportunities to reduce costs. 👉 Check it out: https://lnkd.in/eth6RRE8 #MultiverseComputing #Luminary #EnterpriseAI #AgenticAI