Local models, runtimes and interfaces that fit on owned hardware.
Salvatore Sanfilippo's engine is deliberately narrow: DeepSeek, GLM and Qwen flash models on Macs, DGX Sparks and Strix Halo boxes, with SSD streaming for hardware that is almost big enough.
An English-German mixture of experts, 78B total parameters and 3B active, Apache 2.0 with a vLLM plugin: built for regulated sectors and benchmarked against models with four times the active size.
TypeSafe's decision model charges $0.042 per million input tokens with free output; a parody post and two open rewrites show the trick runs on any GGUF model, and a coin-flip test shows where the numbers get shaky.
Anthropic and OpenAI shipped cheaper flagships on Tuesday, and Anthropic's own announcement says Opus 5.5 is the first release since the pacing essay.
Raspberry Pi's walkthrough runs Cactus's Needle 2 on a Pi 5's CPU: 'turn the LED on' becomes code in 78 milliseconds, with no AI HAT and no cloud call in the loop.
fusion-runtime runs speech recognition, an LLM and speech synthesis in one process, and its benchmark section publishes the number that makes it look worse and explains why the stages do not add up.
