CLI
mach-generate, mach-check, mach-models, and mach-archs.
The package installs one command per task. Every command that takes a checkpoint
accepts either a HuggingFace repo id or a local pack directory, plus --revision and
--cache-dir (default cache: ~/.cache/mach). mach-serve is covered in
Serving.
mach-generate
One-shot generation:
mach-generate SyzygyResearch/Mach-1-Additive-35B \
--prompt "Explain the Doppler effect in two sentences." \
--max-tokens 256When the tokenizer ships a chat template, the prompt is wrapped as a single-turn user
message automatically; pass --raw-prompt for base-model-style verbatim completion.
| Flag | Default | Description |
|---|---|---|
--prompt | required | The prompt text |
--max-tokens | 64 | Generation length cap |
--temperature | 0.0 | 0.0 is greedy decoding |
--top-p / --top-k / --min-p | 1.0 / 0 / 0.0 | Sampling filters |
--seed | none | Seed the RNG for reproducible sampling |
--repetition-penalty | none | Multiplicative penalty (>1.0 discourages repeats) |
--presence-penalty / --frequency-penalty | 0.0 | OpenAI-style additive penalties |
--raw-prompt | off | Skip the chat template |
--prompt-cache | none | Reuse and persist prompt KV state across invocations (.safetensors path) |
mach-check
Preflight a checkpoint before paying for a load. No weights are downloaded:
mach-check SyzygyResearch/Mach-1-Additive-35BReports whether the checkpoint is a servable Mach-1 pack, which serving paths it can
run, a memory verdict (fits / tight / wont_fit) against your machine's unified
memory, and the exact serve command to run next.
| Flag | Description |
|---|---|
--json | Machine-readable output |
--draft-dir | Also probe a draft checkpoint for the speculative-decoding path |
--context-length | Context length for the KV memory estimate |
--local-only | Never touch the network |
--all-archs | Print the support matrix for every registered architecture |
mach-models
List locally cached checkpoints with size and support status:
mach-modelsPass --json for machine-readable output.
mach-archs
Print the supported-architecture registry (--json for the machine-readable form).
The current registry serves the qwen3_5_moe family, the base architecture of the
Mach-1 packs.