Skip to main content

Gradium

Category: Voice / Speech AI

Gradium develops audio language models (ALMs) that enable ultra-low latency, realistic voice AI interactions at scale Gradium was founded in 2025. The company is led by Neil Zeghidour. Based in Paris, France. Team size: 8-20. Total funding raised: $70M. Latest round: Seed, $70.0M, December 2025, led by FirstMark Capital. Key investors include FirstMark Capital, Eurazeo, DST Global Partners, Eric Schmidt, Xavier Niel, Korelya Capital, Amplify Partners, Yann LeCun, Rodolphe Saadé.

Founded
2025
Headquarters
Paris, France
Team size
8-20
Total funding
$70M

Value proposition

Delivers near-instantaneous, expressive, multilingual voice AI with superior accuracy, low latency (sub-200ms), and scalability, outperforming traditional LLMs in speech tasks

Products and solutions

Audio language models for speech recognition, synthesis, and classification, Multilingual support (English, French, German, Spanish, Portuguese), Cloud API with per-minute billing, Edge SDK planned for Q1 2026

Unique value

Spun out from nonprofit AI lab Kyutai; assembled team of top researchers from Google DeepMind, Meta FAIR, Google Brain, and Jane Street; uses natural language supervision on audio-text data for superior voice understanding and generation; focuses on sub-200ms latency for real-time interactions

Target customer

Developers and enterprises building voice-enabled AI applications

Industries served

Gaming, Customer care, Language learning, Healthcare, AI agents, Metaverse experiences, Global voice assistants

Technology advantage

ALMs trained on paired audio-text datasets enable ultra-realistic expressivity, accurate transcription, and sub-200ms latency at scale; commercializes Kyutai's frontier research for B2B deployment; cloud API already operational with edge SDK planned

How they differentiate

Gradium specializes in ultra-low latency audio language models (ALMs) for real-time, multilingual voice AI with superior accuracy, expressiveness, and conversational flow, outperforming general LLMs in speech tasks; spun from Kyutai lab's research like Moshi; currently operational with cloud API

Main competitors

ElevenLabs, OpenAI, Anthropic, Mistral

Key partnerships

Ongoing collaboration with Kyutai for access to latest generative audio research, Plans to open product centers in Berlin and New York

Major milestones

Spun out from Kyutai AI lab (September 2025), Raised $70M seed round (December 2025), Public launch of platform with first ALMs supporting 5 languages, Cloud API currently operational, Plans for Berlin and New York product centers

Growth metrics

Cloud API operational with per-minute billing; plans to expand team to ~80 people within the year; edge SDK planned for Q1 2026

Market positioning

Early-stage leader in low-latency, realistic voice AI for developers building agents, entertainment, and enterprise apps; positioned against crowded field of LLM voice add-ons and specialized synthesizers with focus on scalable, natural interactions and sub-200ms latency

Geographic focus

Europe (Paris, France-based) with initial multilingual emphasis on European languages; global developer/enterprise reach with plans for Berlin and New York product centers

Patents and IP

None publicly disclosed

About Neil Zeghidour

Former Staff Research Scientist at Google DeepMind and Meta FAIR; Founding member and Chief Modeling Officer at Kyutai

Official website: