Skip to content

Retrieve & verify models

simulo models lists and downloads the checkpoints a job actually created, with the download’s integrity checked before anything is written to disk. The training starter uses ResumableCheckpoint: it creates latest.pt for resuming and may create best.pt when it finds a better policy. A training job without checkpoints can have no models to list.

Terminal window
simulo models
NAME KIND SIZE SHA256 CREATED MODEL ID
latest.pt latest 2148728 b7e410c92a55 2026-07-10T18:05:02Z mdl_swift-falcon-3nqk8n-0001
best.pt best 2148728 3f9c2ab81d04 2026-07-10T18:04:18Z mdl_swift-falcon-3nqk8n-0002

MODEL ID has the shape mdl_<adjective>-<noun>-<suffix>-<counter> — the same friendly stem as the model’s job, with a per-job counter distinguishing its checkpoints.

No job id → the most recent job. simulo models <job-id> lists a specific one; a job id prefix (4+ characters) works here exactly like everywhere else in the CLI, for the legacy UUID form some jobs from before friendly identifiers still carry — a friendly id is matched by exact equality only and must be given in full.

Terminal window
simulo models <job-id> best.pt
Saved best.pt (best) -> /home/you/best.pt (sha256 verified)

The CLI downloads the bytes and checks them against the record’s digest_sha256 before writing anything — a corrupted or tampered download fails the command and leaves no file behind, rather than silently handing you a bad checkpoint.

Terminal window
simulo models <job-id> best.pt # download from a specific job
simulo models --all -o ./models/ # every model for the job, into a directory
simulo models <job-id> best.pt --info # print the metadata record as JSON, without downloading
Terminal window
simulo models <job-id> best.pt --info | grep digest_sha256
sha256sum best.pt

Both digests should match — nothing but coreutils needed to confirm the CLI told the truth.

That check proves the bytes arrived intact. It says nothing about whether the policy inside them runs anywhere but Simulo, which is a different question with a different answer: Export a policy to ONNX converts the checkpoint to a portable bundle and ships a verifier that replays known-good inference on your own machine, with no Simulo, simulator, CUDA, or GPU involved.

Keep training past this checkpoint: Continue training. Score it over multiple rounds and record a rollout: Evaluate & roll out.