Local models.
Native execution.

Four of six planned pairs passed the frozen host guards. Two pairs were excluded for external CPU activity and system pageouts. Whole-campaign timing_accepted remains false. These are preliminary local results; quality equivalence and broader production performance remain unqualified.

Rates include prefill, generation, IPC/readout and cleanup. Loading and warmup are excluded. Values are geometric means of within-load medians; each table cell covers 20 requests across four accepted loads. The overall gain gives equal weight to the 24 paired workload/load ratios. Gains use unrounded paired results and can differ from ratios of rounded rates. Bars compare engines within each workload.

Token hashes matched across engines on 6/10 fixtures; decoded text matched on 5/10. Both engines passed only sorting among four strict sanity checks. Arithmetic, prime and Python failed exact-reply checks; the retained receipts cannot distinguish semantic failures from formatting failures.

This measured direct workers at concurrency one. HTTP serving, concurrent requests, time to first token, power and other models or Macs were not measured here. Per-process GPU activity was unavailable; nominal thermal state does not prove equal clock frequencies.

The pinned comparison used Stark 045d0ca6, MLX-LM 72be8c955, MLX/MLX-Metal 0.32.4, context 8192, greedy sampling and fresh request state. Pair-level leads ranged from 6.09% to 7.95%; these are observed ranges, not confidence intervals.

The stock pilot used unchanged engine implementations, normal setup controls and external measurement adapters. Stark, MLX-LM, TensorFold and vLLM Metal used the same MLX affine 4-bit/group-64 checkpoint; Stark used its normal native repack. llama.cpp used the existing Unsloth Q4_K_M export, whose source-weight equivalence to the MLX derivative was not verified.

Every engine ran at concurrency one. vLLM's separate concurrency-eight diagnostic is excluded from these comparisons. Context was 8192 with greedy decoding, thinking disabled and retained/prefix caches disabled through stock controls. Interfaces differed: Stark and MLX-LM used private pipes; the others used loopback HTTP. Missing cache counters remain unknown.

Values are medians of output tokens divided by complete-request seconds. Stark, MLX-LM, vLLM Metal and TensorFold default cover four accepted loads and twelve observations per workload. TensorFold with drafting off and the separately corrected llama.cpp repeat cover three accepted loads and nine observations; their first blocks were excluded for CPU activity. Whole-campaign timing acceptance is false. The llama.cpp driver correction and original rejected samples are retained in the report.

Stark, MLX-LM, vLLM Metal and TensorFold passed only sorting among four exact-answer sanity checks. llama.cpp passed arithmetic, prime and sorting but failed Python. This small screen does not establish generation quality equivalence. Continuous GPU, thermal and power isolation were not measured.

The same pinned JevK5 checkpoint was replayed over 231 public cases and twelve synthetic controls, using BF16 native and GGUF representations. Each decision used context 8192, fresh state, full-vocabulary log probabilities and candidate temperature 1.22. The complete-request timer includes context construction, prefill, normalization, cleanup and transport; loading and warmups are excluded.

Two Stark loads and all four llama.cpp loads passed the timing screens, giving 486 and 972 accepted measured requests respectively. Stark's other two loads were excluded for CPU interference. Whole-campaign timing acceptance is false; medians are descriptive and are not a paired speedup estimate. Input buckets are 0–255, 256–1023 and 1024–8191 tokens.

Public accuracy was 59/231 for both engines. They agreed on 241/243 choices and were individually repeatable across loads, but were not numerically identical. The twelve synthetic controls do not offset the low public accuracy. The sealed replay inputs remain private; public aggregates alone cannot reconstruct that exact replay.