Open models, local AI, self-hosting, and the hardware tradeoffs behind frontier-class open source.
We talked to Cornell researchers, NVIDIA, priced DGX Station at about $100k, and chased the first GLM-5.2 numbers. The question is still memory behavior past 252GB HBM.
GLM 5.2 local requirements depend on the artifact: full 753B, low-bit full-model GGUF, NVFP4, or REAP 504B. Here are the hardware paths people have actually reported.
Ornith-1.0 is drawing attention as an open agentic coding model, but the release is more nuanced than the benchmark chart. Here's what is available, why people are skeptical, and how to test it fairly.
Kimi K2.7 Code lands at roughly one-seventh the output price of Claude Opus 4.8 — $3.50 vs $25 per million tokens — and it's a lot closer on quality than that gap suggests. We pointed Cyrus at it and recorded what happened, plus how opencode support is bringing OpenRouter models to Cyrus.
