Tarsier2-7b-0115

by omni-research

SOTA video description — detailed, temporally-aligned captions that outperform GPT-4o

45Kdl/month

7Bparams

HuggingFace Use in Pipeline

Identifiers

Model ID

omni-research/Tarsier2-7b-0115

Feature URI

mixpeek://video_extractor@v1/omni_tarsier2_7b_v1

Overview

Tarsier2 generates highly detailed, temporally-aligned video descriptions. It achieves state-of-the-art across 16 video understanding benchmarks spanning captioning, QA, grounding, and hallucination detection — outperforming GPT-4o and Gemini 1.5 Pro on video description quality.

For video RAG, detailed description quality is critical: the richer the textual representation of video content, the better text-based retrieval performs. Tarsier2 produces the kind of dense, accurate descriptions that make video truly searchable.

Architecture

7B parameter model from ByteDance research. Optimized for generating faithful, temporally-ordered descriptions that minimize hallucination while maximizing detail density.

Mixpeek SDK Integration

mixpeek.ingest.from_url(
    url="s3://media/interview.mp4",
    collection="video_library",
    feature_extractors=[{
        "type": "caption",
        "model": "mixpeek://video_extractor@v1/omni_tarsier2_7b_v1"
    }]
)