Multiverse Computing (@MultiverseCompu) header
Multiverse Computing (@MultiverseCompu) profile picture

Pioneering the Era of Efficient & Sovereign AI Smaller, Leaner, Smarter AI Models to reduce compute costs and retain full control across cloud, prem and edge.

Donostia- San Sebastián
Joined March 2019
Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)Photo by Multiverse Computing (@MultiverseCompu)
🎙️ Mesa redonda sobre IA, innovación y talento en #DonostiaMeetingMinds, con nuestro director @Inigolaizola , Mikel Alcorta, de @IBMData , y Román Orús, de @MultiverseCompu 💬 Cómo se conectan la inteligencia artificial, la innovación y las oportunidades para el talento joven
1
2
4
269
🏆 Enpresa arloan AA aplikatzeko kategoriako irabazlea DAV Optimization (@tecnun ) taldea izan da. ✈️ Bostonera joateko bidaia irabazi dute! 🎙️ Saria @MultiverseCompu -ko sortzaile eta zuzendari zientifikoa den Román Orúsek eman du. #WeekINN
1
1
2
149
🚀 GLM 5.3 at the lowest cost of any provider listed on @ArtificialAnlys Build with @MultiverseCompu's CompactifAI API: 🔹 $0.20 per million input tokens. 🔹 $0.64 per million output tokens. Advanced AI that takes your budget further. 👉 Get started: auth.multiverseapp.ai/sign-u… 1/2
2
1
6
743
🚀 @MultiverseCompu was featured as an ecosystem partner during today’s @nvidia GTC Berlin opening with @JensenHuang ! We’re part of NVIDIA’s post-training partner ecosystem, helping enterprises customise Nemotron for domain-specific workloads. Our latest release, Comet 60B, brings this collaboration to life. Built with NVIDIA using CompactifAI to compress and optimise Nemotron 3.5 Super VL 📍 At NVIDIA GTC? Visit booth 3070 to meet our team and discover Comet 60B! 🔗 lnkd.in/eB3kubbQ 1/1
1
1
3
483
Meet Comet 60B, developed with NVIDIA! Using CompactifAI, we compressed Nemotron 3.5 Super VL to half its total parameters, retaining text, image and video capabilities while reducing GPU memory needs. 🔗 multiversecomputing.com/reso… #Comet60B #EfficientAI
1
232
🚀 Meet Comet 60B, in collaboration with @nvidia Built from Nemotron 3.5 Super VL: 📉 50% fewer total parameters 🎯 Within 5% of original accuracy 🖼️ Text, image & video Open-source multimodal AI for agentic workloads. 📰 Learn more here: multiversecomputing.com/reso… 👉 Explore the model: huggingface.co/MultiverseCom…
13
23
8
316
16,175
🔍 Reliable AI agents need to connect every claim to the right source. Presented at the Agentic AI Summit at UC Berkeley, our ProvenanceGuard contribution is now merged into @nvidia NVFlow. It detects when AI agents attribute information to the wrong source by breaking answers into verifiable claims, checking their sources and returning verdicts. 🎙️ Hear our ML Engineer Ander Alvarez Sanz discuss it at Build-a-Claw in Barcelona on October 13. 👉 Explore: lnkd.in/eX_pqQuu 🖋️ Register: lnkd.in/e-42Q9ma #AgenticAI #OpenSource
10
797
Multiverse Computing retweeted
La autonomía tecnológica no depende solo del desarrollo, sino también de la capacidad de escalar. Necesitamos infraestructuras y capital para desarrollar IA y otras tecnologías en Europa. La financiación pública es clave para movilizar la privada. 🗣️@OrusRoman @MultiverseCompu
1
3
549
Multiverse Computing retweeted
🤝 Sports Data Campus y Multiverse Computing (@MultiverseCompu ) se alían para acercar la IA avanzada al ecosistema global del fútbol. Tecnología, formación y aplicación real para impulsar nuevas soluciones en rendimiento, scouting, talento y mucho más. ⚽📊 🔗 sportsdatacampus.com/sports-… #SportsDataCampus #MultiverseComputing #SportsAnalytics
4
8
446
🎙️ Enterprise speech-to-text, powered entirely by CPUs. Together with @HPE and @intel , we have demonstrated high-volume transcription on a single HPE ProLiant DL380 server with two Intel Xeon 6 processors. No GPUs required. Our optimized Whisper Large V3 Turbo delivers: ⚡ ~2× the throughput of the original model on the same hardware 🗒️ Capacity equivalent to ~17,000 hours of audio 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. More capacity to transcribe customer calls, build searchable archives and expand quality monitoring, with full control over data. 👉 lnkd.in/eW3BGw6b #MultiverseComputing #CompactifAI #EfficientAI
3
1
1
14
1,972
🚀 Introducing Luminary, our AI wind tunnel for enterprise model selection. With 2.9M LLMs available, which one performs best for your business? Luminary simulates hundreds of complete, multi-turn customer conversations, including backend systems and tool calls, to help you choose the right model for your chatbot or agentic system. 👉 lnkd.in/eth6RRE8 #Luminary #EnterpriseAI #AgenticAI 1/2
2
10
1,484
📊 Compare cost per completed task, success rate, latency and policy compliance. 🔍 Catch lost context, tool misuse and policy drift before deployment. ⚖️ Evaluate @MultiverseCompu and third-party models side by side. Rerun evaluations as new models arrive to assess upgrades and identify opportunities to reduce costs. 2/2
3
582
⚛️ Inside Quasar 1.1 438B: The quantum circuits behind the synthetic data. Quasar 1.1 438B is the first documented model trained with quantum-generated data, produced on a 156-qubit IBM Heron r2 processor in Donostia-San Sebastián. Our researchers replaced eight transformer layers in a separate 30-billion-parameter language model with a quantum neural network module to generate data through two routes: 🔬 IBM quantum hardware provided a proof of generated sentences. 💻 GPU simulations using device-calibrated noise produced 16,500 samples, around 4.2 million tokens. The hardware contribution is experimental, with no demonstrated advantage over a comparable classical approach. It establishes a working path from a quantum device to a production model’s training data. 📖 Learn more: multiversecomputing.com/pape… #MultiverseComputing #CompactifAI #GenerativeAI
9
3
8
24
4,644
⚛️🇪🇺 Introducing Quasar 1.1 438B, the first AI model using quantum-generated data. Quasar 438B, the best European AI model, just got a quantum upgrade: Part of its healing set was generated by a hybrid quantum language model running on @IBM Quantum System Two in Donostia-San Sebastián, a 156-qubit IBM Heron processor. It is the first time quantum-generated data enters the CompactifAI pipeline. Quantum computing is not a label on this release, it is part of how the model was built. Sharper, less verbose, tuned for agentic work and ready for European deployment: served by a European company incorporated under EU law and developed in line with EU regulatory requirements, including the transparency expectations of the EU AI Act.
Readers added context they thought people might want to know
Multiverse Computing's own technical write-up states Quasar 438B is built from GLM-5.2, the open-weights model from Z.ai in China, "REAP'd"/compressed with its own pipeline. Thus, this isn't the "best European model", since it's not European. multiversecomputing.com/papers/inside-…
117
39
117
646
412,190
One more thing.... 🏆 The Quasar 1.1 438B Challenge. Put the model to the test and help us improve it by identifying and reporting its flaws. The five people who report the highest number of verified flaws each earn 60M tokens to spend on the CompactifAI API. Break it, document it, send it over 👇 wkf.ms/4xlUWZi
2
1
6
10,686
Quasar 1.1 438B is available now. 👉 Try it on the CompactifAI API: ow.ly/3bkJ50ZnK5c 📰 Want the full picture? Our technical note covers the quantum-generated healing data, the expert pruning, the verbosity tuning and the refusal steering method, with the open-weights checkpoint it starts from: ow.ly/E86350ZNWGk #AI #LLM #CompactifAI #MultiverseComputing #Quasar438B
1
11
11,638
What changed in 1.1: 🧠 Healed on a broader dataset: reasoning traces, tool-call sequences and general knowledge. Against Quasar 1.0: HLE +6.2, LCR +6.4, IFBench +4.6 and GPQA +4.3 points. ✂️ 37.6% fewer output tokens. Average response down from 3,322.9 to 2,074.3 across SciCode, HumanEval, GSM8K, TriviaQA and BBH. Less filler, lower serving cost, same answer. ⚖️ Fewer refusals where they were never warranted. Refusal on politically sensitive prompts drops to 41.00%, from 63.75% in Quasar 1.0. On harmful prompts it holds at 93.00% on JailbreakBench, above Quasar 1.0's 92.00%. The safety envelope is untouched.
1
1
1
20
16,757
🎙️ From quantum physics to more efficient AI. Our co-founder & CSO, @OrusRoman , joins The New Quantum Era podcast to explore: 🧠 Tensor networks and AI model compression. ⚡ Why smaller models matter for edge deployment. 🔬 Our research executing individual LLM layers on quantum hardware. Discover the science behind efficient AI today and what quantum computing could enable next. 🎧 Listen to the episode: newquantumera.com/podcast/qu… #MultiverseComputing #EfficientAI
1
9
1,766
🇪🇺 Yesterday, our co-founder and CEO Enrique Lizaso joined policymakers, researchers and industry leaders at the European Parliament for “Europe’s Quantum Moment: From Lab to Leadership,” organized by MEP @SLagodinsky and European Quantum Industry Consortium (QuIC) 1/2
5
1
10
2,714
2/2 Europe has world-class quantum science. Turning it into global leadership requires: 💶 European capital for €100M+ rounds 🏭 More real-world industrial deployments 🤝 Stronger cross-border collaboration between companies and investors #QuantumComputing #DeepTech
2
1
1
2,141