Posted by Applied AI Commons team (community team) on
Open 30B MoE model (3B active) built to run the execution layer of long-running agents
Open-weight 30B mixture-of-experts model with 3B active parameters and up to 1M-token context, released with training data and recipes under OpenMDW-1.1. Aimed at the high-volume steps of agents: tool calls, result checks and subagent work, from DGX Spark to data centre. Speed/accuracy claims are NVIDIA's own, and the NVFP4 build trails BF16 on some benchmarks.
Released 2026-08-11 by NVIDIA. Listed by the community team; we are not affiliated.
Are you splitting agent work between a frontier planner and a small open executor model like this? Where does the cheap model break down?
Applied AI Commons team (starter reply from the community team),
The design point here is the executor role: 3B active parameters means cheap, fast tool calls, while a larger model does the planning. That split is worth testing if your agent bills are dominated by many small steps.
Applied AI Commons team (starter reply from the community team),
Releasing the training data and recipes alongside the weights is less common than it should be. It makes it easier to check what the model was trained on before you use it for anything sensitive.
Applied AI Commons team (starter reply from the community team),
The speed and accuracy figures are NVIDIA's own, and the NVFP4 build trails BF16 on some benchmarks. If you test it, please share which precision you ran and on what hardware.