0.14.0 — Measured decode gains, two experimental workspaces, and a much larger catalog

Four measured speed-ups land on the product path, not in a micro-benchmark: qualified verification, a qualified 8-bit pairing, a fused kernel and the scheduler each got faster on a named Mac. Computer Use and Share Compute arrive as explicit opt-ins, the Community Benchmark says what it is doing while it runs, serving starts from one command, and the catalog grows to 190 text aliases plus seven more image models.

• Four measured improvements on the product path. On an M3 Ultra: Qwen3.8 27B MTP GDN verification 45.87 → 52.28 tok/s (+14.0% decode, verification sync 4.954 → 4.644 s); Muse-Glimmer 30B 8-bit DFlash 22.1–22.6 → 41.3–46.3 tok/s (2.00× code median, 1.83× chat); Qwen3.8 Flash-Next 4-bit fused GDN +6.35% end-to-end (+26.28% on the isolated layer); and a short request queued behind three long prompts reached its first token in 1.206 s instead of 3.067 s (2.54× faster). The MTP result used one checkpoint, prompt, seed and K=3 draft depth across three alternating runs with matching output hashes. Fused Flash-Next GDN and shortest-validated-tail scheduling stay explicit opt-ins while their hardware coverage grows.
• Computer Use, as an experiment. Enable it under Experimental Features for bounded workflows that plan and execute locally across supported Mac apps — supervised Downloads cleanup, and drafting content before posting it through Safari or Chrome. Permissions and app targets are scoped, retries are bounded, the visual recovery loop is tied to the active server session, and consequential actions stay reviewable. This is an early preview, not unattended general-purpose automation.
• Share Compute, as an experiment. A new workspace lets you offer compatible local capacity to a configured compute pool. Server and Desktop validate model fit and configuration before accepting work, keep service bindings on safe local interfaces, expose live health and lifecycle state, and propagate cancellation back to the local server. It is off by default and does not participate until you opt in.
• The Community Benchmark tells you where it is. Desktop and CLI now show the selected model, the current phase, elapsed time and an ETA while a run is in progress. Rows share one definition of median decode throughput and TTFT, record run conditions and cached model metadata, distinguish local results from shared ones, and preview contributor identity before upload. Results stay local unless you explicitly share them.
• Ten more text aliases. g9v3-39a5b-4bit is a 21.3 GiB mixed-precision MoE for 32 GB+ Macs, its reasoning probe decoding at 85 tok/s on an M3 Ultra. Granite 4.2 arrives as six aliases (30B/8B/3B × 4-bit/8-bit), all through the same 38-check server journey, measured from 20.8 tok/s (30B 8-bit) to 175.1 tok/s (3B 4-bit). minicpm5-2b-4bit is a compact 1.3 GiB option that scored 24/31 in 7.17 s against 26/31 in 32.54 s for the 4B comparison — faster, with the quality tradeoff documented. neohorse-9b-4bit measured 44.5 tok/s but did not beat the default across the full quality suite, so it ships available rather than selected.
• Seven more image models. FLUX.1 schnell, Stable Diffusion 3.5 Large, SDXL Base, Bonsai Image 4B 2-bit, HiDream O1 Dev and Qwen Image Edit join the catalog. From the release qualification matrix: FLUX.1 schnell at 1024×1024 in 23.36 s with 9.46 GiB peak, Bonsai 4B 2-bit at 512×512 in 5.39 s with 6.52 GiB peak, SDXL Base at 1024×1024 in 16.46 s after construction. These are single-machine measurements; the readiness UI shows the memory tier before a costly load.
• Serving starts from one command. rapid-mlx start picks a fitting cached model, starts the canonical server, waits for readiness and configures a supported agent profile — --dry-run previews every mutation. rapid-mlx service installs and operates a least-privileged headless macOS service with status, logs, restart, uninstall and rollback. rapid-mlx doctor adds lifecycle, import, cache, configuration and service diagnostics with verified repairs. Desktop can unload an idle resident model in one click while keeping the selection, and it refuses to interrupt active chat, image, audio, video or dictation work.
• An optimization we turned off. Qwen3.5 4B MTP is now off by default: release qualification measured single-stream decode 25–37% slower on M2 Pro and M3 Ultra with no concurrency gain. It remains available explicitly. Telemetry now honours RAPID_MLX_TELEMETRY=0, DO_NOT_TRACK=1 and common CI markers in both the engine and the Desktop app.