Unverified Apple M5 Ultra Local LLM Benchmarks Surface on oMLX

A Reddit post highlighted unconfirmed LLM inference benchmarks published on the oMLX website for an unannounced Apple M5 Ultra processor. Running a 4-bit quantized 27B model (Qwen 3.8 27B q4) at an 8k context window without multi-token prediction, the listed metrics show a generation throughput of 50 tokens/second and prompt processing at 1,800 tokens/second. Although unverified, these numbers offer an early speculative look at the potential local AI performance of future high-end Apple Silicon hardware. If accurate, such high prompt processing and generation throughput on a 27B model demonstrate how unified memory architectures could continue to streamline running heavy local LLMs on Macs. The reported benchmark highlights prompt processing speeds of 1,800 tokens/second and generation throughput of 50 tokens/second achieved without Multi-Token Prediction (MTP). However, because Apple has not officially announced or detailed an M5 Ultra chip, these figures lack technical verification and must be treated as speculative third-party claims.