Neon's team says it used 1,300 H200s and months of experimental data to further train an open-source model and apply reinforcement learning. The resulting model, Neon, surpassed GPT-6 Astra on the team's own analysis benchmark, it claims. Its initial materials-science focus includes superconductors, magnets and semiconductor materials. A team member also shared a demo described as Neon investigating a real sample from a superconductor lab—hypothesizing about crystal structure, reasoning about synthesis conditions and iterating with tools until it finds a physically realistic solution.