AI news for Nov 8, 2026

The developer behind llama-modchip published version 2.4.0, an Apache-2.0 toolchain for running mixture-of-experts models entirely in system RAM on machines without GPUs. It also covers Nemotron-3.5-30B-A3B and Qwen3-30B-A3B, and ships prebuilt Linux x86-64 binaries, a Windows hybrid GPU/CPU launcher, and macOS setup, with commands for chat, benchmarking, quantization and a local Claude Code session. The custom quantized tensor format requires the project's patched runtime rather than stock llama.cpp, Ollama or LM Studio.