Marlin-2B: this tiny 2B model pinpoints moments in video

Locating an exact moment in a 30-minute clip has always been a painful task. Most vision-language models can write a decent summary, but they completely fall apart when you ask for precise second-level timestamps.

AI lab NemoStation just open-sourced Marlin-2B to fix that exact headache. Fine-tuned from Qwen 3.5 2B Base and released under the Apache 2.0 license, this tiny 2B model runs easily on consumer hardware.

Pinpointing what happened and when

Marlin-2B was designed for a single, practical job: extracting structured facts from raw video. It answers two specific questions: what happened, and exactly when did it happen?

You can feed it a clip and receive a dense narrative broken down by start and end timestamps. Or you can query a specific action, such as searching for a shootout, and it will return the exact time boundaries where that event took place.

That makes it an immediate asset for video editors scrubbing for b-roll, security teams reviewing surveillance footage, content moderators, and developers labeling massive video datasets without paying enterprise API bills.

Despite its compact footprint, Marlin-2B punches well above its weight class. On TimeLens-Bench, an evaluation that bypasses text-based LLM judges, it ranks fifth among open-source models. On combined averages, it beats GPT-5 by 3.9 mIoU despite being roughly 3.5 times smaller, and rivals models like Gemini 2.5 Flash in dense captioning accuracy.

Auditing the broken benchmarks

The most fascinating part of the release is what the team uncovered during development. When NemoStation attempted to evaluate Marlin-2B against industry standards like CaReBench and DREAM-1K, the scores looked suspiciously off.

They discovered that the video-AI field has been relying on text-only judges that never actually inspect the video. The judge simply places the candidate output next to the benchmark's ground-truth text and checks for matching words.

This created absurd penalties for accurate models. In one test clip showing a yellow creature, Marlin-2B correctly described it as a yellow humanoid. But because the human annotator had omitted the word yellow from the ground truth, the text judge ruled it an unsupported hallucination.

NemoStation audited both benchmarks using a video-aware multi-frame judge. The results were startling: ground-truth errors appeared in 70.2% of CaReBench clips and 65.8% of DREAM-1K entries. Benchmark baselines were actively misidentifying sleeve lengths, misreading on-screen banner text, and imagining actions that never occurred.

Why it matters

We think Marlin-2B is a masterclass in pragmatic model design. Instead of building another mediocre generalist, NemoStation poured a tight 2B parameter budget directly into the temporal axis.

For independent creators, developers, and small agencies, the model is an instant win. Video comprehension via commercial APIs gets expensive fast and poses data privacy concerns with raw footage. At just two billion parameters, Marlin-2B can sit quietly on a local GPU or modern laptop, tagging thousands of video clips offline for free.

More broadly, this release exposes a massive blind spot across multimodal AI research. When standard benchmarks are riddled with flawed ground truth and evaluated by blind text judges, the leaderboard race becomes theater. By calling out broken evaluations and delivering a capable local model, NemoStation gives practitioners a tool that actually works in production.