Command-line reference

The dynalm command line downloads models (pull), chats (run, chat), serves an OpenAI-compatible API (serve), stops a server (stop), inspects models (inspect, models), diagnoses the machine (doctor) and measures performance (benchmark, autotune). Run dynalm help or any command without arguments for its options.

Commands at a glance

DynaLM commands
Command	Purpose
`dynalm pull <name\|link>`	Download a model (registry name, Hugging Face link or any `.gguf` URL)
`dynalm run <model> [-p TEXT]`	One answer with `-p`; an interactive chat without it
`dynalm chat <model>`	Interactive chat (same as `run` without `-p`)
`dynalm serve <model>`	OpenAI-compatible HTTP server (one model per server)
`dynalm stop`	Ask a running server to finish requests and exit
`dynalm models`	List local models; `--available` lists registry names
`dynalm rm <model>`	Delete GGUF files, partial downloads or Hugging Face folders
`dynalm inspect <model>`	Architecture, size, quantization, RAM estimate and a support check
`dynalm doctor [--json]`	System report: CPU features, backends, memory, kernel self-test
`dynalm benchmark <model>`	Load test: throughput and P50–P99 latency
`dynalm autotune [--run <model>]`	Show kernel selection; tune the CPU thread count
`dynalm config [show\|path\|init]`	Effective configuration and its sources
`dynalm version`	Versions, build type, compiler and kernels

Model references

Wherever a command takes <model>, you can pass a registry name (qwen3:4b, llama3.2:3b), an alias (qwen3, llama:3b), a GGUF file path (./models/qwen3-4b.gguf) or a Hugging Face model folder (./models/st/qwen2.5-0.5b-instruct).

Coming from Ollama

Ollama to DynaLM commands
Ollama	DynaLM
`ollama pull qwen3:4b`	`dynalm pull qwen3:4b`
`ollama run qwen3:4b`	`dynalm run qwen3:4b`
`ollama run qwen3:4b "question"`	`dynalm run qwen3:4b -p "question"`
`ollama list`	`dynalm models`
`ollama rm qwen3:4b`	`dynalm models rm qwen3:4b`
`ollama show qwen3:4b`	`dynalm inspect qwen3:4b`
`ollama serve`	`dynalm serve qwen3:4b` (one model per server)
`ollama stop`	`dynalm stop`

`dynalm run` options

run options
Option	Meaning
`-p, --prompt TEXT`	The prompt (without it, `run` starts a chat)
`-n, --max-tokens N`	Tokens to generate (default 128; chat: 2048)
`--system TEXT`	System message
`--chat` / `--raw`	Use the model’s chat template (default) / send text as is
`--stop TEXT`	Stop when this text appears (repeatable)
`-t, --threads N`	Compute threads (default: physical cores)
`-c, --ctx N`	Context / KV cache size in tokens (default 4096)
`--gpu` / `--cpu`, `--backend ...`	Device selection; see backends
`--mode M`	`auto`, `performance`, `exact` or `safe`
`--temp`, `--top-k`, `--top-p`, `--min-p`, `--seed`	Sampling (default for `run`: greedy)
`--spec ngram` / `--spec DRAFT.gguf`	Speculative decoding; with greedy sampling every kept token is the model’s own choice

Chat commands

Chat commands
Command	Effect
`/bye`, `/exit`, `/quit`	Leave (also Ctrl+D)
`/clear`	Forget the conversation
`/system TEXT`	Set the system message and start over
`/think on` / `/think off`	Qwen3 reasoning on or off
`/set temp 0.7`, `/set top_p 0.9`, `/set seed 42`	Change sampling
`/stats on`	Print speed after each reply
`"""`	Start and end a multi-line message
Ctrl+C	Stop the current reply

The model stays loaded and earlier turns are reused rather than reprocessed, so follow-up messages start quickly. When the conversation outgrows the context (-c), the oldest exchanges are dropped.

`dynalm benchmark` options

benchmark options
Option	Meaning
`--concurrency LIST`	For example `1,4,16` (default `1,4`)
`--prompt LIST`	Prompt lengths in tokens (default `128,512`)
`--output LIST`	Output lengths (default `128`)
`--requests N`	Requests per point (default 2 × concurrency, at least 4)
`--url http://HOST:PORT`	Benchmark a running server (DynaLM, llama.cpp, vLLM, Ollama…)
`--out FILE`	Append results as JSON lines

Example

dynalm benchmark qwen2.5:1.5b --concurrency 1,4,16

Server options are listed in OpenAI-compatible API.

Global options

-q prints errors only, -v adds debug detail and --log-level L accepts quiet, normal, verbose, debug or trace. To hide [INFO] lines in a chat: dynalm --log-level warn chat ....

Frequently asked questions

Does `run` download the model automatically?

Yes. run, serve and benchmark download a named model on first use if it is not cached.

How do I download a gated model such as official Llama?

Accept the license on Hugging Face, then set HF_TOKEN before dynalm pull.

Source: docs/cli.md (opens in a new tab)