NemoStation
2B video model — dense captions and plain-English timestamp search
Marlin-2B is a 2-billion-parameter video VLM fine-tuned on Qwen3.5-2B, purpose-built for two tasks developers actually need: dense captioning (what happened + when, structured with second-precise timestamps) and natural-language temporal grounding (query with plain English, get exact start/end timestamps back). At ~4–5 GB, it fits a 6 GB consumer GPU, beats Qwen2.5-VL-7B by +6.4 mIoU on TimeLens-Bench, and matches Gemini-2.0-Flash on temporal grounding. **Step-by-step:** 1) `pip install transformers`. 2) `from transformers import AutoModelForCausalLM, AutoProcessor`. 3) Load: `AutoModelForCausalLM.from_pretrained('NemoStation/Marlin-2B', trust_remote_code=True)`. 4) Call `.caption(video_path)` for dense timestamps or `.find(video_path, 'your query')` for NL search. **Best for:** video indexing, security/event detection, meeting search, sports highlights.
Query any video in plain English and get back precise start/end timestamps.
Free (open-source on Hugging Face) · ~4–5 GB VRAM