Command line

mixturekit build CONFIG.yaml [--output DIR] [--dtype bfloat16|float16|float32] [--no-streaming]
                             [--push-to-hub ORG/NAME] [--private]
mixturekit train CONFIG.yaml
mixturekit visualize MODEL [--text TEXT] [--html FILE] [--device cuda|mps|cpu] [--port 8501]
mixturekit card MODEL [--repo-id ORG/NAME]
mixturekit push MODEL ORG/NAME [--private]
mixturekit export MODEL OUTPUT
mixturekit benchmark CONFIG.yaml

build

Merges the experts described in the YAML file (see Configuration) and saves the configuration it was built from, with the datatype used, as mixturekit_config.json. --push-to-hub uploads the result.

train

Trains the routers or stitch layers (see Training). Needs mixturekit[train].

visualize

Opens the routing app, or with --text or --html prints the share of every expert and writes an HTML view (see Routing visualizer). Needs mixturekit[vis] for the app.

card

Writes the model card (README.md) of a merged model.

push

Uploads a merged model with its card to the Hugging Face Hub. Training state (checkpoints, optimizer) and the local build configuration are not uploaded; the card holds the build configuration with the local paths removed. --private refuses to upload into an existing public repository.

export

Writes a traditional MoE of Llama, Mistral or Qwen3 experts as a native Mixtral or Qwen3-MoE model (see Export to Mixtral, Qwen3-MoE and GGUF).

benchmark

Scores merged models and their experts with lm-evaluation-harness (see Benchmarks). Needs mixturekit[bench].