In Hemres Agent, it is possible to assign agents specifically to
AUXILIARY TASKS models in addition to the main model (deepseek v4 pro).
I have now selected the following options—which are not permitted—based on cost considerations.
What do you think? @Stll0@oneitonitram
This model was just recently released (in early June 2026) by NVIDIA. It’s a massive 550-billion-parameter model (with 55B active parameters per token via MoE) featuring a hybrid Mamba/Transformer architecture. Most importantly: It’s specifically trained for agents, deep logical reasoning, and orchestration.
Unlike “Owl-Alpha” (which is an anonymous stealth model with privacy concerns and high latency), Nemotron-3-Ultra comes from NVIDIA, is extremely high-performing (high throughput), and reliable.
Since this model is free, has a context size of one million tokens, and (unlike Owl-Alpha) offers massively intelligent reasoning for agent flows, it once again completely upends my previous “Smart Economy” table in an extremely positive way. I can now replace expensive OpenCode Go flat-rate models with the free NVIDIA model!
Thanks for the tip.
What I won’t replace: Title Gen / Approval / Compression: Groq: Llama 3.1 8B (These tasks require millisecond response times for the UI. Although Nemotron is fast, nothing beats Groq’s LPU chips for these micro-tasks).
