The 80B Qwen3-Coder-Next on a 16 GB Mac, about 13 tok/s, fully offline. A mixture-of-experts inference engine for Apple silicon.
Coachwhip Crew
Four CrewAI agents (analyst, architect, developer, tester) work a coding task through Coachwhip's local OpenAI door, on one screen, every word streaming. A runner (no AI) executes the developer's code; a judge runs a fixed set of checks the agents never see.
crew_app.py— the server and the crew.python crew_app.py, then open http://127.0.0.1:8072/ and press Run. Needscrewai(pip install crewai) and, for the runner and judge, Node withnpx tsx.crew_app.html— the page.task_lru.txt— the task. Change it for your own work.judge_lru.ts— the judge: eight executed checks for the LRU task. Write your own for your task; it imports./main, the developer's code.record_app.py— records a run to MP4 at 1280x720 (needsplaywrightand its Chromium).
Env: COACHWHIP_BASE (default http://127.0.0.1:8090/v1), CREW_PORT (8072), CREW_MAX_TOKENS (3000). For a cloud door: CREW_CLOUD=1, CREW_CLOUD_BASE, CREW_CLOUD_MODEL, CREW_CLOUD_KEY_FILE (a file holding the key; never the key itself), CREW_PRICE_IN / CREW_PRICE_OUT in dollars per million tokens for the cost chip.
Nothing leaves your machine on the local door: the page listens on 127.0.0.1 only.