Tried NVIDIA's free Nemotron 3 Ultra as a coding agent model in OpenCode. What I found.

A developer tested NVIDIA's free Nemotron 3 Ultra (550B MoE, 55B active, 1M context) as a coding agent model within OpenCode. The experiment aimed to assess its performance without incurring API costs. The developer is now seeking community input on whether other free or open-weight models have successfully handled tool calls in agent setups, or if paid, larger models remain necessary for practical applications. This highlights a common challenge in AI development: balancing cost-effectiveness with performance for specialized tasks like coding assistance.

This report offers a first-hand account of using NVIDIA's Nemotron 3 Ultra as a coding agent, unlike many discussions that focus on theoretical capabilities or benchmarks.