AI Briefing
KO

Marlin-2B Video Structuring VLM Released

·2026.05.19 18:52

Key point

Marlin, a 2B open-source VLM based on Qwen3.5-VL, has been released.

Details

Marlin-2B is a 2B open-source VLM based on Qwen3.5-VL, tailored for video understanding across two practical tasks.

  • caption(): Outputs Scene descriptions and Events structured with second-level timestamps
  • find(): Returns natural language queries as (start, end) segments of the video

The developers stated that rather than being a general-purpose chat-style VLM, it was narrowly designed for the functionality needed in video editing and structuring pipelines. vLLM inference is also provided.

In terms of performance, they explained it is best-in-class among comparable open models on DREAM-1K, CaReBench, and TimeLens-Bench, and also emphasized cost efficiency compared to Gemini-2.5-Flash.

The training recipe and new video captioning/grounding benchmarks are planned to be released in the future.

open-sourcevision-language-modelvllmvideo-captioningtemporal-groundingqwen3-5-vl
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

PyTorchKRvLLM Introduces Hardware-Agnostic Layer Architecture to Support Diverse AcceleratorsPyTorchKRvLLM rope_scaling Override Causes 36% Model Configuration MismatchRedditTextCLF Releases Calibration-Free TQ Quantization Method and Open-Source Quant FactoryRedditTextCLF Releases Calibration-Free 4-bit Quantization ToolHacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesHacker NewsJevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection