GLM-5.2
Z.ai's MIT-licensed open-weight MoE (~744B total, ~40B active), tuned for long-horizon agentic coding with a 1M-token context. Server-class, but Unsloth's 2-bit dynamic GGUF (UD-IQ2_M, ~239GB) fits a 256GB+ Mac Studio for a private, slow, tireless local agent.
Alternatives
5 workflows use GLM-5.2
Run GLM-5.2 for the bulk, escalate the hard turns to Opus 4.8
Wire a cost-routing config that sends most work to cheap hosted GLM-5.2 and only the hardest turns to Opus 4.8, instead of paying Opus prices for everything.
Route through a gateway with a tested open-weights fallback
Keep model access from being a single point of failure: route through an OpenAI-compatible gateway and pin a fallback that is open-weights and has actually been tested, so a pulled or deprecated model is a two-minute config change, not a lost week.
Wire GLM-5.2 into Hermes: valid route, 64k-context check, no key in config
Validate a Hermes Agent config that runs GLM-5.2 through a real provider route (direct Z.AI or OpenRouter), clears Hermes's 64k minimum context, and keeps the API key out of config.yaml, before you start a session.
Serve GLM-5.1 yourself for long-horizon agentic coding (vLLM)
Stand up the MIT-licensed GLM-5.1 FP8 checkpoint as an OpenAI-compatible endpoint for long agentic runs, validated serve config + endpoint.
Run GLM-5.2 fully local on a Mac Studio and drive it with Hermes
Serve GLM-5.2's 2-bit GGUF on a Mac Studio over an OpenAI-compatible endpoint, point Hermes at it as a custom provider, and hand it long hands-off agentic jobs.