lobes

A machine running lobes serves a small fleet: an always-warm text primary, a multimodal gear, tiny embedding and reranking gears, audio in and out — fronted by one gateway that routes each request by its model field. Callers address a lobe by role — cortex, senses, embedder, reranker — and never need to know which checkpoint answers.

Every verb speaks --json, results go to stdout and diagnostics to stderr, and every write verb is a dry run until you say --apply. Agents call CLIs in loops; safe-by-default is not a nicety here, it is the contract.

The tour in four verbs: whoami tells you which machine and which model. status tells you the configured model and container health. switch changes the served model — showing you the full resolved config before touching anything. And capabilities tells you, machine by machine, which lobes are ready.

A model never runs bare. lobes resolves its serve config through three layers — the machine profile, the workload purpose, and the model's own catalog entry — and any explicit flag overrides them all.

That last clause is the point: the profiles are shipped defaults, not fences. New machine? Add a profile. Different workload? Pass --purpose, or override a single knob with a flag. Open for extension, open for override.

Here is the part worth leaning in for. When this page first went up, this section described a direction. As of lobes 0.42.0 it describes a shipped choice: a deployment shape — which of the six lobes a box hosts. machine-as-brain, the default, keeps the whole brain on one box, exactly as before. The mesh shapes give each box the lobes it is best at: spark-lobe keeps the Qwen cortex and drops Gemma senses; thor-lobe keeps Gemma senses and drops the Qwen cortex.

The reason is arithmetic. Co-residency taxes every lobe: sharing the Spark, cortex is trimmed to 131072 context at 0.30 GPU util so senses can fit alongside. Move the box to its shape — lobes init --shape spark-lobe, dry-run by default, --apply to commit — and cortex gets its machine back: 262144 context, the checkpoint's full native window, at 0.44 util. The mirror move on Thor serves senses at 131072, four times its 32768 co-resident trim. Both shapes were validated live on the physical boxes on 2026-07-14; the numbers are deployed env values, not estimates.

And a dropped lobe is dropped honestly. capabilities flags it, /v1/models omits it, and a request for it returns 404 role_infeasible — never a silent reroute to a different model. Each box tells the truth about what it hosts, so composing boxes into one brain stays addressing, not archaeology.

Every deployment claim above is measured, and the harness ships in the box. lobes benchmark measures decode throughput and prefill latency — its request shape follows the configured --purpose, so the numbers track what the machine is actually tuned for. lobes assess runs correctness probes against the served model, in markdown ready for a per-model doc.

The setup is not a lab secret: the profiles are a pure data module, the compose templates are packaged in the repo, and each runtime model has a doc recording how it ran, its live test results, and its caveats. Clone it, run the same two verbs, compare.