# Rapid-MLX 0.15.4 headline models — M4 Pro 48 GB (2026-10-02)

Raw data behind the homepage "Hot models" / Performance numbers and the
`/models/` hot-model pages measured on this machine.

- Machine: Mac mini, Apple M4 Pro, 48 GB unified memory (see `env.txt` for the
  full `sysctl`/version dump; macOS 26.5.1).
- Engine: `rapid-mlx==0.15.4` from PyPI in a fresh venv. `gemma-4-26b-4bit` and
  `bonsai2-27b-2bit` ran after `rapid-mlx[vision]==0.15.4` was installed into
  the same venv (they need the optional vision runtime); everything else ran on
  the slim base install. No other flags.
- Harness: the published /compare harness (`bench_compare.py`, seed 20260926,
  temperature 0, streaming, `include_usage`), scenarios:
  - **a** — single-stream decode, median of 3 × 512-token generations.
  - **b** — cold TTFT at a ~2,000-token and a ~16,000-token prompt, median of 3
    (`ttft_2k_median_s` / `ttft_16k_median_s` in `summary.json`).
  - **d** — 4 concurrent streams × 2 rounds, aggregate decode
    (`aggregate_4stream_median_tps`).
- `summary.json` is the reviewed roll-up; each model directory holds the
  per-scenario raw request JSON (`a.json`, `b.json`, `d.json`), one-line
  medians (`*.summary`), the `/v1/models` record (`models.json`),
  `/v1/status` snapshot after the run (`status.json`), and the server's own
  precheck (`precheck.txt`, `info.txt`). Server logs are omitted (they contain
  local paths). Decode rates are computed from engine-reported usage token
  counts, never client estimates.
- Peak memory: server-reported `/v1/status` `metal.peak_memory_gb` (with the
  MLX active figure alongside); `peak_rss_gb` is the fallback ps sample of the
  server PID tree. The two are labelled separately wherever they appear.
- `qwen3.8-27b-tensorfold` was also measured the same way; its profile is an
  experimental opt-in and those numbers are not part of this published set.

Every page citing this run links here and labels the chip, engine version and
date (M4 Pro 48 GB · rapid-mlx 0.15.4 · 2026-10-02).