Technical preview
NativeRig measures speed, memory and a real tool call on your machine, activates only what passes, and rolls back if an upgrade fails.
Measured, not estimated. Time to first token, throughput and resident memory come from a run on your machine, not from a catalog.
Gated on a real tool call. A model that cannot read a file through OpenCode never becomes your default.
Rolled back, not left broken. A failed candidate or upgrade restores the last configuration that passed.
How NativeRig works
Ollama, Docker and LM Studio will get a model running. NativeRig checks whether the exact model, runtime and client combination actually works on your Mac before it becomes your default, and keeps the last passing setup recoverable.
NativeRig reads your chip, memory and free storage and compares them to the qualified list. If your Mac is not on it yet, setup stops before any download and you can join the waitlist.
NativeRig verifies the model bytes, tunes context, GPU offload, threads and memory reserve, then measures first-token time, throughput and resident memory and requires a real OpenCode tool call.
A failed candidate never replaces your working setup. A passing one opens in managed OpenCode without touching your existing global or project settings, and a failed upgrade rolls back.
Download
This is a pre-release build. Only configurations that passed the complete coding, tool, offline and rollback gates are listed. Windows and Linux builds exist for lab use and are not offered here until they pass the same gates.
| Mac | Model | Runtime | First token | Generation | Resident | Tool call | Measured |
|---|---|---|---|---|---|---|---|
| Apple M5 Max · 128 GB | Qwen3-Coder-Next Q8_0 | llama.cpp b9637 · Metal | 134.39 ms | 70.55 tok/s | 84.4 GB | passed | 2026-09-06 |
Mac preview
Apple silicon · macOS 14+
SHA-256a1f56e9c8158…Unsigned pre-release test build. Your operating system will warn before installation. Automatic updates are disabled.
Limited live validation. Review the release notes and remaining qualification work before testing.
First setup needs internet access and downloads the qualified model, about 80 GB for the current configuration. NativeRig shows the exact files and sizes and asks first. After setup, the coding profile runs offline by default.
For current testers
The feedback form is on a separate page. Bring the CSV NativeRig saved, then report whether installation and first launch worked.