Community experiments push large-model inference onto limited local hardware

Community reports continue to test MoE tiering and constrained hardware, but the 284B/64GB claim and Kimi K3’s 0.85 tokens/s 1-bit Optane run lack strong independent validation. A 120B-MoE-on-M1-Max report using a 16 GiB budget stresses matched-weight comparisons, memory pressure, and correctness pitfalls, but its single-Mac scope and unreleased engine limit actionability. Recent Kimi K3.1 Pro deployment coverage is B300-specific, untested by its author, and largely duplicates an earlier guide; hardware guides should be treated cautiously when speed figures are estimates.