Model

DeepSeek-VL2

OPEN

DeepSeek · DeepSeek-VL family · released Dec 13, 2024

MoE VLM family (Tiny/Small/Base) with dynamic image tiling; Base variant scores strongly on OCRBench (811) and DocVQA (93.3) among open VLMs of its era.

ReasoningCodingVisionFunction callingTool useAgentic

953.6

Elo · unrated

Parameters

27B

Active params

4.5B (MoE)

Context

4K tokens

Architecture

MoE vision-language transformer (Base variant), 27B total / 4.5B active — DeepSeekMoE+MLA language backbone + SigLIP-SO400M vision encoder with dynamic tiling (family also ships Tiny: 3B/1.0B-active, Small: 16B/2.8B-active)

License

DeepSeek Model License (code: MIT)

Languages

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision81.4%#34

best: Molmo 72B · 96.3%

ChartQAVision86.0%#19

best: MiniMax-VL-01 · 91.7%

DocVQAVision93.3%#17

best: Qwen2-VL-72B · 96.5%

MathVisionVision17.3%#21

best: Seed 2.1 Pro · 92.6%

MathVistaVision62.8%#39

best: Seed 2.1 Pro · 90.7%

MMBench (Chinese)Vision79.6%#6

best: ERNIE 4.5 VL 424B-A47B · 90.9%

MMBench (English)Vision83.1%#14

best: Qwen3.5-397B-A17B · 93.7%

MMBenchVision79.2%#2

best: Phi-3.5-vision (4.2B) · 81.9%

MMMUVision51.1%#92

best: Claude Fable 5 · 89.3%

OCRBenchVision811#10

best: InternVL3-78B · 906

RefCOCO+Vision91.2%#1

best: this model · 91.2%

TextVQAVision84.2%#5

best: Molmo 2 8B · 85.7%

Run it locally

VRAM @ Q4

14 GB

VRAM @ FP16

54 GB

Fits on (Q4)

RTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

No public RTX 4090 llama.cpp benchmark found; no confirmed GGUF/MLX conversion of the vision-language model

Quantizations

—

DeepSeek-VL family

Elo progression across releases

DeepSeek-VL 7BMar 202439.6 DeepSeek-VL2Dec 2024953.6

Resources

HFdeepseek-ai/deepseek-vl2 GHdeepseek-ai/DeepSeek-VL2

API price weights · each benchmark row carries its own source badge (see methodology)