nemostation 0.1.22

Find moments in your videos. Agent-first CLI for Marlin-2B: dense captioning + temporal grounding, local or hosted.

The command-line tool for Marlin-2B — a 2B video VLM for the two questions you actually ask a video: what is happening, and when. Runs free and local on Apple Silicon — no API key, no Hugging Face account.

• marlin caption → a Scene description + a <start>–<end> event timeline
• marlin find → the single start → end span where your query happens

Install

Apple Silicon (M-series Mac) only for now. NVIDIA / other platforms are coming as a separate optimized build.

uv tool install nemostation      # 1. install   (or: pipx install nemostation)
marlin setup                     # 2. set up    (sign in, build engine, download weights)
marlin caption clip.mp4          # describe what's in a video
marlin find clip.mp4 "a deer crossing"   # locate when it happens → start → end

Two commands and you're done — setup does everything: a one-time browser sign-in (two questions, then Google), builds the local MLX engine, and downloads the weights. After it finishes, caption and find just work. The 8-bit weights are public — nothing gated, no API key. Add --json to any command for parseable output. (ffmpeg is optional — only for windowing videos >2 min.)

The engine stays warm between calls so responses are fast. To shut it down and free the RAM (~16 GB): marlin stop. It auto-starts again on the next call.

What it produces

marlin caption "video.mp4" — what's in it	marlin find "video.mp4", "gunfight" — when it happens

Each call runs one model pass on one bounded clip (~2 min at 2 fps) — the same contract as the inference server. Clips are auto-downscaled to the model's ~200K-pixel budget before inference (faster, far less memory, no accuracy loss) — tune with --max-pixels (lower on weak machines) or --full-res to opt out. For longer videos, window with ffmpeg and loop.

Why Marlin

At 2B params it's the strongest open model in its weight class on dense captioning (DREAM-1K, CaReBench) and natural-language temporal grounding (TimeLens-Bench) — competitive with Gemini-2.5 at a fraction of the cost. See the benchmarks on the model card.

Use it from an agent

marlin skills install        # → .claude/skills/ + .agents/skills/

Installs the video-understanding skill so Claude Code / Codex use marlin as "eyes on a video" — clip-length and single-find limits baked in. Every verb honors --json (stdout parseable, progress on stderr).

Contributing

Marlin is meant to be extended — and adding a skill is the easiest way in. A skill is a folder under skills/ with a SKILL.md that teaches an agent to use caption / find for one job — clip scoring, b-roll search, highlight reels, footage catalogs, whatever you build.

skills/
  video-understanding/SKILL.md   # ships today — the reference
  your-skill/SKILL.md            # ← add yours

Add one:

1. Copy the format from video-understanding/SKILL.md — frontmatter (name, description, requires.bins) + a short recipe.
2. Keep it honest about the limits (one bounded clip per call; find returns one span).
3. Open a PR. New skill ideas, issues, and docs fixes are all welcome too.

Hack on the CLI:

git clone https://github.com/shu-bamma/marlin-cli
cd marlin-cli
uv tool install --editable .     # or: pip install -e .
pytest                           # contract tests

New verbs, engine support, and bug fixes are all fair game — open an issue to chat about anything bigger. Licensed under Apache-2.0.

Links

• Try it live → vlm.nemostation.com
• Model card + benchmarks → huggingface.co/NemoStation/Marlin-2B
• Team / custom fine-tuning → nemostation.com