Use cases
What this archive is for, and which machine to use.
Security research and red-team prep
GLM-5.3-Flash or Flash-Next on one M5; abliterated builds when a stock model refuses legitimate defensive questions (malware triage, exploit write-ups for patched CVEs, phishing-awareness samples).
Where: M5 ×1 (Flash-Next IQ4_XS / alesha IQ3_S), M5 ×2 for DeepSeek V4.1 IQ1_M
Keep abliterated models offline and off public endpoints. Authorised targets only.
Local coding agent
Serve Flash-Next IQ4_XS with llama-server and point an OpenAI-compatible coding agent at it. Only 6B params are active per token, so it is the fastest large model here (theoretical ceiling ≈67 tok/s on one M5). Qwen3.8-27B Q4 on the MacBook Pro for travel.
Where: M5 ×1, MacBook Pro
Code never leaves the LAN. Qwen Community License allows internal commercial use.
Private document Q&A
Open WebUI + llama-server + your own documents. Long native context (262K on Qwen3.8, 1M advertised on GLM-5.3/DeepSeek V4.1) helps, but KV cache grows with context; start at 32K.
Where: M5 ×1
Client documents stay on hardware you own.
Content for Remnant apps
Qwen-Image 2.1 / MOSS-Image for concept art, Wan 2.2 or HunyuanVideo for short clips, ACE-Step for background music on SnapDog/CyberSim promos.
Where: M5 ×1 (image/music), M5 with offload (Wan)
Check licenses: Qwen-Image 2.1 and YuE2 are non-commercial; MOSS-Image (Apache 2.0), Wan 2.2 (Apache 2.0) and ACE-Step (MIT) are commercial-OK.
Voice: transcripts, narration, voice notes
Whisper turbo for meeting/podcast transcripts; Qwen3-TTS (Apache 2.0) for narration with your own cloned voice; Fish S2 Pro for research comparisons only.
Where: Every machine except the NAS for TTS; Whisper even runs on the Mac mini
Only clone voices with consent.
Vision: screenshot and UI review
All four LLM families are image-text-to-text; load the mmproj file with llama-server to ask about App Store screenshots, error dialogs or diagrams.
Where: M5 ×1, MacBook Pro (27B)
Use the mmproj that ships beside each GGUF.
Quantization lab
Original safetensors are archived so you can make your own GGUFs or try new quant methods (GSQ-RCO, Mirai) without re-downloading 1.2 TB.
Where: Any box with disk; M5 to test results
Originals are the source of truth; keep the archive offline-safe.