Models

Every tier answers "what do you run on" plainly. CroweLM names the product layer: knowledge-base grounding, the cultivation and vision personas, tier metering, the apps, and the routing between engines. The engine column names the model that does the inference. Crowe Logic does not train those engines; what it built is the layer on top, and one model of its own at the bottom of the table.

The table

Grouped by the plan that unlocks the row. A higher plan includes every row below it. Ask any tier what it runs on and it gives the same answer as this table.

The seven rows from GPT-5.5 to Claude Opus 4.8 are the frontier picks: the engines by their own names, with no CroweLM identity and no CroweLM base prompt. Each introduces itself as what it is. They also run on your own keys on any plan; crowe init --node scaffolds the env file. Kimi K3 is deployed, but on 2026-08-31 its upstream was refusing calls with a rate limit; a pick that cannot answer falls through the chain like any throttled tier. Engines were verified end to end on 2026-08-31 and this table is updated the day an engine under a tier changes.

Effort rungs

CroweLM, Zenith, Depth, Flash and Coder each run on one deployment at three rungs: normal, high and max. /effort normal|high|max moves within the family you are on; /effort alone reports the current rung. What the knob does depends on the engine: Claude rows set the model's own effort level (medium, high, max), Depth raises Grok's reasoning effort from low to medium to high with 8k, 16k and 32k output budgets, Flash and Coder raise the output budget. A rung is also addressable directly, as in /model zenith-high or /model crowelm-max. The plan ladder follows the rungs: CroweLM normal on Personal, high on Pro, max on Max; Flash normal on Free, high and max on Personal; every rung of Zenith, Depth and Coder on Pro.

Picking a tier

/model prints the table with the engine each tier runs on and the active one marked. /model 2 or /model kimi switches by number or name; /model resolve <alias> shows what an alias maps to right now. /model auto hands the choice to CroweLM Auto, which reads each turn and routes it to a family by task class; a bare crowe starts there.

crowe route "prompt" shows the decision without invoking a model: the intent, the tier the router would pick and the engine behind it. --json emits the same as JSON.

$ crowe route "refactor this python module to remove the duplicated retry loop"

 Prompt          refactor this python module to remove the duplicated retry loop
 Intent          code
 Confidence      0.86 (ok)
 Selected        CroweLM Coder (crowelm-coder)
 Runs on         gpt-5.6-sol by OpenAI
 Tier type       code
 Runtime params  temperature=0.2, top_p=0.95, max_tokens=6144
 Reason          intent=code; primary=CroweLM Coder; 3 fallback(s); 0 companion(s)

Two tiers at once

/dual on streams two tiers side by side on the same prompt; the default pair is CroweLM and CroweLM Zenith, and /dual <a> <b> picks your own. If the right pane is on a tier the plan cannot reach, it falls back to one it can. /dual synth on adds a synthesis pass after both panes finish, in one of three modes: merge fuses the strongest claims, judge picks the better answer, diff shows where they disagree. /dual off returns to one tier. Dual mode sums the credits of both sides.