DeepSeek-VL2
Mixture-of-experts vision-language model family in tiny, small, and full sizes
About
DeepSeek-VL2 is DeepSeek's second-generation vision-language family, released in December 2024 with three sizes: Tiny (1B activated), Small (2.8B activated), and the full model (4.5B activated). The mixture-of-experts design keeps inference cheap while handling visual question answering, OCR, document and table understanding, charts, and visual grounding. The team reports an OCRBench score of 834 and 93.3% on DocVQA, ahead of GPT-4o on those document-heavy tasks. It is genuinely useful for invoice, screenshot, and dashboard reading rather than creative image work. Weights are public under DeepSeek's custom model license, which is not an OSI-approved license, so commercial users should read it first.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
OCRBench 834, DocVQA 93.3% (DeepSeek-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Release date 2024-12-13 and three size variants verified on the project's GitHub release log; OCRBench 834 and DocVQA 93.3% are DeepSeek-reported figures from the model paper, labeled as such; Weights public under DeepSeek's custom model license (not OSI-approved); commercial-use caveat captured in description.