NVIDIA GTC Berlin Brings Agent Demos, Mistral Partnership, and New Coding Model

At GTC Berlin 2026, NVIDIA demonstrated its Agent Toolkit across three industry sectors: Lila Sciences using agents to design potential B-cell lymphoma therapies guided by an automated lab, Siemens testing new-product production plans in a digital twin of its Erlangen factory before changes reach the floor, and Dassault Systèmes using engineering agents to identify design problems and refine vehicle designs through simulation. NVIDIA attributed the agents' cross-domain capability to its CUDA-X accelerated libraries. On the same day, Mistral and NVIDIA announced a collaboration on AI for industrial engineering, covering faster simulation, open benchmarks and datasets for physics AI, and evaluations for agentic AI in computer-aided engineering and electronic design automation. Also released at GTC Berlin was Recursive Flint, a 40-billion-parameter mixture-of-experts coding model from Recursive Superintelligence with only 2.5 billion active parameters, achieving over 220 tokens per second on RTX Spark. Recursive said the model scores at the top of agentic coding benchmarks including SWE-Bench Pro v2 and Terminal-Bench 2.1, and that its Eureka Machine automated research system boosted Flint's generation speed by 55 percent without modifying model weights. Flint's open weights are available on Hugging Face under the Apache 2.0 license, with post-training support via NeMo Automodel and NeMo RL.