• Sources: primary, discussion
  • Summary: NVIDIA says Nemotron 3.5 Bolt 74B-A9B uses quantization-aware training to fit 2-bit weights into 24GB and targets local agents on RTX Spark machines shipping 2026-10-16. It will run at launch in llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, and NVIDIA says the training data, weights, and recipes are open.
  • Why it matters: A fully open agent-oriented MoE that fits on a single 24GB GPU gives local coding-agent setups another candidate to evaluate, pending independent benchmarks.
  • Follow-up: Weights and model card on 2026-10-15, then independent quality results for the 2-bit build.