Reviewing Nvidia Nemotron 3 Super and Releasing Our Own Model!

TL;DR

Nvidia's 120B Nemotron 3 Super and the community-built Kappa model both run locally, enabling powerful multi-agent AI orchestration through the open-source Turnstone platform.

Key Points

• 1.Nemotron 3 Super specs: 120B parameter mixture-of-experts model with only 12B parameters active at once; runs locally on Nvidia Spark/GB10 hardware with 128GB VRAM in NVFP4 or FP8 format with 131K context window
• 2.Kappa model release: A fine-tune of GPT-OSS 20B with D&D-style character alignment, strong long-context recall, and tool calling; runs in as little as 10GB VRAM (MXFP4), available on HuggingFace
• 3.Turnstone platform: An open-source, community-built multi-agent AI orchestration tool (by Patrick) that connects local and cloud models, supports tool calling, and is deployable via `docker compose up`
• 4.Context engineering over prompt engineering: Modern AI workflows are shifting to feeding models rich context (files, documentation, man pages) rather than crafting clever prompts, especially for multi-agent setups
• 5.The "car wash problem": Nemotron 3 Super defaults to recommending walking due to alignment training overriding logical reasoning; Kappa's D&D alignment makes it more likely to push back and reason correctly
• 6.Kappa vs. Nemotron on reasoning: In some alignment and tool-calling scenarios, the 20B Kappa model outperforms the 120B Nemotron because it glazes less, pushes back on bad ideas, and reasons more honestly