The simulator runs an always-hot router and one cold expert. Each graph is Qwen2.5-0.5B
Choice, MatMulNBits q4, 453.87 MiB.
0.991Router · n=115
0.838Shell-safety · n=105
1.000Function router · n=98
Planted gold, slim Choice ONNX. Router labels none / shell_safety / function_router.
Shell-safety labels allow / warn / block. Function router labels none / shell_exec /
read_file / write_file / web_search / ask_user. The function-router panel can share
phrasing with its train set; the export check is zero letter disagreements against the
full head.
Qwen letter token ids. MiniCPM 54–59 are the legacy shell-safety package.
Package
K
Labels
Token ids
Router
3
none / shell_safety / function_router
32 / 33 / 34
shell_safety
3
allow / warn / block
32 / 33 / 34
function_router
6
none … ask_user
32–37
How it works
One forward on the router returns choice_logits with shape
[batch, 3]. Route and score loads the chosen expert and runs one more
forward. The router stays loaded. One expert stays with it. Switching experts unloads
the previous expert.
What you can run
Router alone, an expert alone, or router then expert on one turn. Evaluate scores the
package in the picker. Route and score always starts at the router.
Limits
Labels are planted gold. Peak accounted size is the router plus one expert, about
908 MiB. The menu scores; it does not set policy.
Default package is the MoJE router. Load, then Evaluate for the router alone. Route and
score runs the router and, when the route is an expert, loads that expert for a second
forward. Each q4 file is 453.87 MiB from
Hugging Face via same-origin /hf.