NVIDIA Nemotron 3.5 Super VL + Bolt for always-on and local agents
At a glance
Open 120B-A12B multimodal MoE (Super VL) for specialized always-on agents with tool use and document/vision understanding, plus Bolt 74B-A9B 24GB for local agentic work on RTX/DGX Spark; weights on Hugging Face/OpenRouter/NIM, Bolt E4B due Oct 23. This workflow is listed under open-weights, multimodal, agent-models, local-inference, edge.
Agents / tools
Nemotron 3.5 Super VL, Nemotron 3.5 Bolt, NVIDIA NeMo, Hugging Face, OpenRouter, NIM
Open-weight 27B decision model answers many choice/noul/score questions about one state in a single noncausal joint request (Decision Index ~68; vision ~77), with a calibrated readout head and optional fast packing path—pattern for cheap agent routing/tool selection.
US lab Reflection open-sourced Beam (501B MoE / 23B active, 1M context, Apache 2.0) for coding and long-horizon agents, with day-0 vLLM/SGLang/Transformers support, Mirror CLI harness, and 40+ inference partners including OpenRouter and Ollama.
EtherPK’s /Agent (Alt+A) opens a read-only agent chat inline in a markdown/outliner PKMS: you pick context scope (block tree, document, links, or whole-graph tools). BYO keys (Anthropic/OpenAI/Gemini/OpenRouter/Ollama); provider calls stay in-browser; only selected turns save back as outline blocks. Solves cold-start chat and lost answers by anchoring agents in durable notes.