hummin¶

A GLM-native terminal coding agent. Local inference is a first-class citizen, not a compatibility mode: colibri, llama.cpp and Ollama on your own hardware, surfaced in one model picker alongside Z.ai's GLM catalog.

export HUMMIN_COLIBRI_INSTANCES="http://nas:9996,http://nas:9998,http://mac:9998"
hummin
# /model -> GLM-5.3-Flash [NAS], Qwen3.8-27B [Mac], your whole fleet, one picker

Get started Run local models

See it in action¶

Ornith 1.5 35B (local GGUF via llama.cpp) working through four upstream bugs in hummin's own source. Watch the footer: live tok/s next to the context bar - 31 tok/s from a model running on a Mac Mini. The full path to this setup is in Local Models.

hummin running Ornith 1.5 35B locally at 31 tok/s

Why hummin¶

GLM first

GLM-5.3, GLM-5.3-Flash and GLM-5.3-highspeed ship in the built-in zai catalog with 1M-token context and mapped reasoning variants. The default model is GLM.
Local fleet

One provider across every OpenAI-compatible server on your LAN: health-probed, serialized per origin, startable straight from the model picker.
Agent teamwork

Sessions message each other across terminals and projects, share a task board, spawn bounded subagents, run scheduled wake-ups.
Safety that bites

Seatbelt/bubblewrap sandboxing, destructive-command advisories, loop and budget guardrails, checkpoints with /rewind.
Memory that persists

Session distillation into lessons plus a self-curating Obsidian-compatible vault with BM25-ranked recall.
Built for slow models

Queue-aware UX, real context windows, idle-timeout tuning and a local-vs-cloud cost split - because a 195GB model streams from disk.

From zero to a local model in the terminal¶

Download

Resumable hf download, stall-proof flags, the right machine for the WAN pipe.
Serve

llama.cpp on Mac (launchd) or Linux (compose); colibri for frontier MoE.
Connect

One env var, /model, done. Fleet start/stop from the picker.
Operate

Service templates, co-run rules, warm caches, measured numbers.