Emmi AI

Emmi AI is an Austrian deep tech company founded in 2024 that develops Large Engineering Models (LEMs) — pre-trained AI systems designed to replace conventional physics solvers with instant, physics-accurate simulation results. The company was acquired by Mistral AI in May 2026, with the acquisition announced as a move to create "the leading AI stack for industrial engineering." Emmi AI's open-source framework, Noether, underpins its model development and is available on GitHub under a non-production license.

What It Is

Emmi AI builds foundation models for the physical world — transformer-based neural networks trained on high-fidelity simulation data that can predict physics outcomes (fluid dynamics, heat transfer, structural mechanics) in real time rather than hours or weeks. The core thesis is that traditional simulation software, largely unchanged since the 1970s, creates bottlenecks across aerospace, automotive, semiconductor, and manufacturing industries. Emmi AI's LEMs are designed to collapse the boundary between CAD design and CAE validation, enabling simultaneous design and simulation.

The Noether Framework

The Noether Framework is Emmi AI's open software stack for building, training, and deploying Engineering AI models. Key characteristics include:

• Modular transformer architecture optimized for physical systems
• Hardware agnostic: runs on CPU, Apple Silicon (MPS), and NVIDIA GPUs
• Multi-GPU and SLURM cluster support for industrial-scale workloads
• Installable via pip as emmiai-noether
• Benchmarks show training the AB-UPT model on ShapeNet-Car takes 5.7 minutes on an NVIDIA H100 vs. 135 minutes on a MacBook Pro M3 Max

The framework is licensed under a Non-Production License (based on Mistral AI's MNPL), meaning research and development use is free but commercial/production deployment requires a separate commercial license.

Production Models

Emmi AI has released two named Large Engineering Models:

• NeuralWing: Validates aircraft wing design in real time, targeting aerospace engineering workflows
• NeuralMould: Described as the first digital engineer for injection moulding, released in February 2026, enabling real-time simulation for plastics manufacturing

Both models are built on the Noether Framework and share unified modeling assumptions, standardized software architecture, and consistent data handling approaches.

Update: Mistral AI Acquisition (May 2026)

In May 2026, Mistral AI announced a definitive agreement to acquire Emmi AI. According to the announcement, the acquisition is intended to strengthen Mistral's position as an AI transformation partner for industrial enterprises and accelerate the development of Physics AI for engineering and manufacturing. Emmi AI's Chief Scientist Johannes Brandstetter — previously at CERN and Microsoft, and co-architect of Aurora (described as the world's first AI foundation model for weather forecasting) — published a post on the rationale for joining Mistral. The Noether Framework's latest release at the time of acquisition was v2026.4.0, published April 29, 2026.

Backing and Team

The company raised a €15M seed round, described on its About page as the largest ever for an Austrian startup, backed by European deep tech investors. The founding team includes Johannes Brandstetter (Chief Scientist), Dennis Just (CEO), and Miks Mikelsons (COO). The Noether repository lists endorsements from research groups at ETH Zurich, JKU Linz, University of Pennsylvania, University of Washington, TU Munich, and Sorbonne University.