NVIDIA (Oct 16) details RTX Spark as a single-machine coding-agent workstation: decision models in llama.cpp/Unsloth choose next actions up to ~200x faster than long reasoning loops; HydraFusion prompt router escalates only when local models aren’t enough; GitHub Copilot gains Windows ML + llama.cpp local backend; Junie Local and Pi optimized. Pair with open Nemotron Bolt 74B-A9B 24GB for on-device agents via Ollama/vLLM/LM Studio.
IBM Granite Switch 4.1 (Oct 15, HF blog): 8B/30B compose chat + Decision aLoRA + Guardian safety aLoRA + Granite-Speech-5.0-Turbo-CTC ASR in ~base VRAM via Activated LoRAs with shared KV cache—swap chat/decision/speech on one endpoint; vLLM plugin and llama.cpp support; community can mix Granite Libraries adapters.
Oct 15 ship: text+image decision model returns calibrated yes/no, choice, or score probabilities in one forward pass—no generation tokens—at $0.042/M input; pitched for defect checks, claim photos, and agent routing gates.