Minimum specs & the three ways to run NHJ¶

← Not-Happy-Jan docs

Not-Happy-Jan scales from "runs on anything, quietly" to "the full Australian bogan call-centre, live." It's all the one install — these tiers are just switches.

Tier	What you get	Needs	Resident RAM
Silent	markers + haptics/displays, no audio	any Mac · `nhj mute`	~nil
Pre-recorded	the cast speaks from the bundled voice bank (ships in the package)	any Mac (`--minimal` install)	~nil (no model loaded)
Full experience ⭐	live cloned voices + the dynamic bogan brain	Apple Silicon M1+, 16 GB rec.	~3.2 GB while warm

Tier 0 — Silent / text-only¶

Run nhj mute (or use a host with no audio output). NHJ still fires its [vibes:] events to whatever you've wired up — the haptic mouse, a Ulanzi pixel display, an ESP32 bell — with no sound. The lightest possible footprint; runs on any Mac. (Note: TTS_ENGINE=none alone is not silent — it disables only live synthesis, and the bundled bank still speaks.)

Tier 1 — Pre-recorded samples (the "she'll be right" tier)¶

A --minimal install runs NHJ entirely from the bundled voice bank — a set of pre-rendered character lines that ships inside the package. Out of the box, with no model and no downloads, Jan, Bazza and Karren speak for each feedback state with nothing resident — the mixer just plays a WAV. You get the cloned voices and the whole call-centre flow on any Mac.

The trade-off: a finite repertoire. The lines come from a fixed bank, so they repeat, and you don't get fresh, dial-scaled slang or the full swearing ladder. Great for older Macs, RAM-tight laptops, or anyone who wants the vibe without the heavyweight bits. (nhj build-bank lets you pre-render your own lines on top, if you've added a custom voice.)

Tier 2 — The full experience ⭐ (where the magic happens)¶

This is what Not-Happy-Jan is for. Turn on the two local models and the whole thing comes alive:

Local Qwen3-TTS (~2.2 GB) — every line is synthesised live, in the character's own cloned voice. Nothing is canned.
ocker-bogan-nano (~0.94 GB) — NHJ's own custom swear-bear-bogan LLM, crafted from the highest-quality drunken Macca's reviews and the unfiltered depths of Australian Reddit. It writes a fresh line in character every single time, scaling from polite reception all the way to full-tilt, ockerism-11, bleeped-or-not swearing.

Together they deliver the full Australian bogan call-centre experience: Jan picks up and mangles your variable names, Bazza escalates when the build goes warm, and Karren has a complete meltdown when it all falls over — every time in a live voice, saying something you've never heard before.

Specs for the full experience¶

	Requirement
Chip	Apple Silicon (M1 or later) — TTS and the LLM run on Metal / MLX
RAM	8 GB works · 16 GB recommended (~3.2 GB resident while both daemons are warm)
Disk	~5 GB — TTS model (~2.2 GB) + ocker-bogan-nano (~0.94 GB) + media (~250 MB)
Network	once, to fetch the models + media — then 100% local & private

About that ~3.2 GB — and how to get it back¶

Models are downloaded once into ~/Library/Caches (about 5 GB) and served two ways. On-demand (default): the first vibe starts the server, it serves the line holding about 3 GB of RAM, then after about ten minutes idle it unloads and frees the RAM, repeating each cycle. Persistent (full profile): the daemons stay always warm at about 3.2 GB resident for sub-second replies with no cold start.

Warm daemons give sub-second replies (no cold-start), at ~2.1 GB (TTS) + ~1.1 GB (LLM) resident while loaded — nothing on a Mac Studio, noticeable on a 16 GB laptop.

The default install uses on-demand model loading; bash install.sh --full opts into the always-warm persistent daemons instead. You can also switch at runtime:

nhj servers on-demand     # start servers on the first vibe, unload after idle
nhj servers persistent    # back to always-warm
nhj servers               # show the current mode + which servers are up

In on-demand mode nothing is resident at rest: the first dynamic/TTS vibe starts the server, and a supervisor reaps it after ~10 min idle (NHJ_SERVER_IDLE), so you only pay the ~3 GB while NHJ is actually talking — at the cost of a one-off cold-start on the first vibe after a quiet spell. The default install sets up the dynamic LLM; set NHJ_DYNAMIC=0 to run TTS-only, and TTS_ENGINE=none to fall back to the bundled bank (which pays nothing).

Switching tiers¶

Want…	Do
Silent	`nhj mute`
Pre-recorded only (the bundled bank)	`bash install.sh --minimal` — no models; the cast speaks from the bundled bank
The full experience (default install)	`bash install.sh` — TTS + the `ocker-bogan-nano` LLM + media, loaded on demand
The full experience (persistent daemons)	`bash install.sh --full` — same, kept warm — see Install profiles

Where the models and media live on disk is covered in Configuration & requirements. See Install profiles for the full breakdown of default · full · minimal.