Murphy R on Mythryl: “Local LLM benchmark: 7 model/host combos, 3 passes each, one agentic…”

Local LLM benchmark: 7 model/host combos, 3 passes each, one agentic task Seven local model/host combinations on an M5 Max with 128 GB, three passes each on the same agentic coding task, averaged: prefill, decode, memory, MTP acceptance, cache hit rate and code quality....

mrmurphy.devLocal LLM benchmark: 7 model/host combos, 3 passes each, one agentic taskSeven local model/host combinations on an M5 Max with 128 GB, three passes each on the same agentic coding task, averaged: prefill, decode, memory, MTP acceptance, cache hit rate and code quality.

15/9/26, 12:48 am0 reposts0 likes

Post