Every basert command and its flags. Run basert <command> --help for the
authoritative, version-specific list.
Download a model into the local hub cache — a pre-converted .base from the
catalog, or a raw HuggingFace repo converted on the fly.
basert pull <id> [options]| Flag | Default | Purpose |
|---|---|---|
<id> |
— | basecompute/<name> (catalog) or org/model (raw HF repo). |
--profile <path> |
— | Override the auto-selected quant profile. |
--target <scheme> |
base-q4 |
Quant scheme for convert-on-pull when no profile applies. |
--revision <ref> |
main |
HF revision / branch / tag. |
--force |
off | Re-download / re-convert even if cached. |
--dry-run |
off | Resolve and print the plan without downloading. |
List models in the local cache.
basert list [--remote] [--json]| Flag | Purpose |
|---|---|
--remote |
Also list catalog models that aren't installed yet. |
--json |
Emit JSON instead of a table. |
Convert a source model (GGUF / HF / MLX) to .base.
basert convert <input> [options]| Flag | Default | Purpose |
|---|---|---|
<input> |
— | Source model path (or a bundle name with --synthetic). |
-o, --output <path> |
<input>.base |
Output .base file. |
--target <scheme> |
— | Quant scheme (or fallback when --profile is set). |
--profile <path> |
— | Per-tensor canonical-quant profile JSON. |
--calibration <file> |
— | UTF-8 calibration text (required for AWQ). |
--calibration-tokens <N> |
— | Number of calibration tokens. |
--awq-mode <mode> |
— | AWQ calibration mode. |
--awq-profile <path> |
— | Precomputed AWQ activation-stats sidecar. |
--synthetic |
off | Generate a dummy bundle (CI/testing). |
See Converting models.
Summarize a .base file's header, tensors, and slots.
basert inspect <input> [--verify-checksums]| Flag | Purpose |
|---|---|
--verify-checksums |
Also verify per-tensor xxhash64 (slow for large models). |
Generate an ed25519 keypair for signing.
basert keygen --output <dir> [--name <name>]| Flag | Default | Purpose |
|---|---|---|
-o, --output <dir> |
— | Output directory. Writes <name>.secret and <name>.pub. |
--name <name> |
baseRT-key |
Base name for the key files. |
Sign an unsigned .base file.
basert sign <input> --output <path> --key <secret> [--key-id <id>]| Flag | Default | Purpose |
|---|---|---|
<input> |
— | Unsigned input .base file. |
-o, --output <path> |
— | Signed output .base file. |
--key <path> |
— | Path to a 32-byte ed25519 secret key. |
--key-id <id> |
baseRT-default |
Human-readable key identifier stored in the file. |
Verify a signed .base file (exits non-zero on tampering).
basert verify <input> --pubkey <path>| Flag | Purpose |
|---|---|
<input> |
Input .base file. |
--pubkey <path> |
Path to a 32-byte ed25519 public key. |
These exec basert-<cmd>. Full flags in their guides.
Selecting a quant variant. For the model-taking commands (serve, chat,
complete, bench, profile), choose which quant build to load with an inline
:variant on the hub id, or the --variant flag:
basert serve basecompute/gemma-4-E2B-it:default-q8
basert serve basecompute/gemma-4-E2B-it --variant default-q8 # equivalentRun basert list to see installed variants (default-q4, default-q8, …). An
uninstalled variant is fetched on demand. The inline :variant wins if both are
given.
OpenAI-compatible HTTP server. See Serving an API.
basert serve --model <path> [--model <path2> ...] [options]Key flags: --host, --port, --api-key, --max-context, --max-tokens,
--kv-bits, --paged-kv, --continuous-batching [N], --prefix-cache,
--rate-limit, --idle-timeout, --request-timeout, --drain-timeout,
--files-dir, --log-file.
Interactive chat. See Chat & completion.
basert chat <model> [options]Key flags: --temp, --top-k, --top-p, --min-p, --repeat-penalty,
--max-tokens, --max-context, --think/--no-think, --kv-bits,
--paged-kv, --max-batch-size.
One-shot completion (text / image / audio).
basert complete <model> --prompt <text> [options]Key flags: --prompt, --system, --chat, --image, --audio,
--max-tokens, --temp, --top-k, --top-p, --repeat-penalty, --kv-bits,
--paged-kv.
Throughput benchmark. See Benchmarking.
basert bench <model> [-p N] [-n N] [-r N] [-w N] [--paged-kv]Audio transcription (Whisper-class models).
basert transcribe <model> <audio> [options]Profile prefill/decode timing for performance investigation.