DeepSeek logo

DeepSeek VL2 Tiny

Multimodal
DeepSeek

An advanced series of large multimodal Mixture-of-Experts (MoE) Vision-Language models that significantly surpasses its predecessor DeepSeek-VL. DeepSeek-VL2 demonstrates superior capabilities across various tasks including but not limited to visual question answering, optical character recognition, document/table/chart understanding, and visual grounding.

Key Specifications

Parameters
3.0B
Context
-
Release Date
December 13, 2024
Average Score
63.1%

Timeline

Key dates in the model's history
Announcement
December 13, 2024
Last Update
July 19, 2025
Today
March 25, 2026

Technical Specifications

Parameters
3.0B
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Multimodal

Working with images and visual data
AI2D
testSelf-reported
71.6%
ChartQA
testSelf-reported
81.0%
DocVQA
testSelf-reported
88.9%
MathVista
testminiSelf-reported
53.6%
MMMU
AI: val valSelf-reported
40.7%

Other Tests

Specialized benchmarks
InfoVQA
testSelf-reported
66.1%
MMBench
testSelf-reported
69.2%
MMBench-V1.1
cn testSelf-reported
68.3%
MME
Standard evaluation AI: methods formation for solutions. should be exclusively and on contextSelf-reported
19.1%
MMStar
Standard evaluation AI: (GPT-4o/Claude/etc.)Self-reported
45.9%
MMT-Bench
Standard evaluation AI: The magic bullet is a model's ability to solve most questions in a benchmark given one try, or more generally, to solve many questions in one go.Self-reported
53.2%
OCRBench
Standard Evaluation Standard evaluation AI: need to evaluate problem and solution. I and her/its solve, on data and mathematicalSelf-reported
80.9%
RealWorldQA
Standard evaluation AI: Translation descriptions model artificial intelligence on Russian language - standard evaluation performance and capabilitiesSelf-reported
64.2%
TextVQA
In deep training and machine training, relates to to evaluation and testing models for verification their efficiency and to This not simply verification accuracy, but also evaluation abilities model data, which she/it not and her/its in real conditions. includes in itself: 1. set data for not at training 2. various scores efficiency 3. on and 4. 5. on to examples 6. Analysis cases, when model gives incorrect In LLM often includes also evaluation by such how: - Accuracy information - and answers - to in various tasks - Quality reasoning helps that model to and that in her can improvements on basisSelf-reported
80.7%

License & Metadata

License
deepseek
Announcement Date
December 13, 2024
Last Updated
July 19, 2025

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.