by NemoStation
2B-parameter open-weights video-language model that writes scene descriptions and timestamped event captions.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Marlin 2B is a video-to-text vision-language model from NemoStation, built on Qwen 3.5 2B. It takes video input and produces scene descriptions and timestamped events, covering video captioning and temporal grounding.
A YouTube creator reported that it matches Gemini 2.5 Flash on captioning tasks, with a download size under 6 GB. This comes from a secondary source and has not been independently verified here. The small size makes it practical for local or low-resource deployment.
Weights are published on Hugging Face under Apache-2.0 in safetensors format for the transformers library. The repository is tagged custom_code, so loading it may require trusting remote code. No website, paper or documentation was found in the provided material.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
2B video language model, based on Qwen 3.5 2B, that produces scene descriptions and timestamped events. Reported as matching Gemini 2.5 Flash on captioning; under 6 GB.
Closest neighbours by meaning, not just by category.
Model · Google
Google's 2B-parameter Gemma 2 instruction-tuned model, repackaged as a text encoder for PixelDiT and PiD workflows in ComfyUI.
Model · SenseTime
Open-weights 27B model from SenseTime that creates short videos by generating and rendering executable code.
Model · NVIDIA
NVIDIA's open multimodal model (120B total, 12B active parameters) for image, video and document understanding.
Model · OpenBMB
Small open-weights multimodal model from OpenBMB for image and video understanding on phones and edge devices.
Model · Black Forest Labs
9B-parameter open-weights image generation and editing model from Black Forest Labs, built for low-latency use.
Research
Research video world model that reasons in a learned discrete "physical language" before rendering video.
