Model

DeepSeek-VL 7B

OPEN

DeepSeek · DeepSeek-VL family · released Mar 8, 2024

First DeepSeek VLM (1.3B/7B); hybrid SigLIP+SAM vision encoder trained jointly with the LLM to preserve language ability.

ReasoningCodingVisionFunction callingTool useAgentic

39.6

Elo · rank #398

Parameters

Active params

7B (dense)

Context

4K tokens

Architecture

Dense 7B transformer (DeepSeek-LLM-7B base) + SigLIP-L/SAM-B hybrid vision encoder (up to 1024x1024)

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

AGIEvalReasoning27.8%#39

best: OLMo 3-Think 32B · 88.2%

GSM8KMath55.0%#125

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning68.4%#131

best: Claude 3 Opus · 95.4%

MathVistaVision36.1%#59

best: Seed 2.1 Pro · 90.7%

MBPPCoding35.2%#91

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMBench (Chinese)Vision72.8%#9

best: ERNIE 4.5 VL 424B-A47B · 90.9%

MMBenchVision73.2%#5

best: Phi-3.5-vision (4.2B) · 81.9%

MMLUKnowledge52.4%#230

best: OpenAI o3 · 92.9%

MMMUVision36.6%#105

best: Claude Fable 5 · 89.3%

OCRBenchVision456#13

best: InternVL3-78B · 906

SEED-BenchVision70.4%#5

best: NAVER HyperCLOVA X SEED Think 32B · 77.9%

Run it locally

VRAM @ Q4

5 GB

VRAM @ FP16

14 GB

Fits on (Q4)

RTX 3060 12GBRTX 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 for the full vision-language model

Quantizations

MLX

DeepSeek-VL family

Elo progression across releases

DeepSeek-VL 7BMar 202439.6 DeepSeek-VL2Dec 2024953.6

Resources

HFdeepseek-ai/deepseek-vl-7b-chat GHdeepseek-ai/DeepSeek-VL

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