Llm Ops

DeepSeek-V4-Flash activates 13B parameters per token but keeps 284B resident. Its experts ship in FP4, and that one choice sets almost the whole memory bill.

Two API settings moved GPT-5.6 Sol from 13.3% to 38.3% on ARC-AGI-3 with no weight change. Why the harness, not the model, decided that benchmark score.

GPT-5.6 Pro is unannounced and every spec is a leak. A builder's checklist to make any frontier launch a non-event: pin model IDs, freeze evals, cap cost, name a fallback.