Model

InternVL2.5-78B

OPEN

OpenGVLab (Shanghai AI Lab) · InternVL 2.5 family · released Dec 5, 2024

First open-source MLLM to exceed 70% on MMMU, rivaling GPT-4o while trained on far fewer tokens than competitors.

ReasoningCodingVisionFunction callingTool useAgentic

1350.1

Elo · rank #147

Parameters

78B

Active params

78B (dense)

Context

32K tokens

Architecture

InternViT-6B vision encoder + Qwen2.5-72B language model (dense composite MLLM)

License

Qwen License (InternViT component MIT; Qwen2.5-72B-Instruct component under Qwen License)

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

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

AI2DVision95.7%#4

best: Molmo 72B · 96.3%

ChartQAVision88.3%#9

best: MiniMax-VL-01 · 91.7%

CharXivVision42.4%#33

best: Muse Spark 1.1 (xhigh) · 88.4%

DocVQAVision95.1%#8

best: Qwen2-VL-72B · 96.5%

MathVisionVision32.2%#18

best: Seed 2.1 Pro · 92.6%

MathVistaVision72.3%#20

best: Seed 2.1 Pro · 90.7%

MMBench (Chinese)Vision88.5%#3

best: ERNIE 4.5 VL 424B-A47B · 90.9%

MMBench (English)Vision88.3%#9

best: Qwen3.5-397B-A17B · 93.7%

MMMU-ProVision48.6%#46

best: Claude Opus 4.7 · 85.5%

MMMUVision70.1%#54

best: Claude Fable 5 · 89.3%

OCRBenchVision854#8

best: InternVL3-78B · 906

OlympiadBenchMath11.6%#7

best: Doubao-1.5-Pro · 59.8%

RefCOCO+Vision90.4%#2

best: DeepSeek-VL2 · 91.2%

RefCOCOVision93.7%#2

best: Qwen3.5-Omni-Plus · 95.0%

TextVQAVision83.4%#7

best: Molmo 2 8B · 85.7%

Video-MMEVision72.1%#16

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

53 GB

VRAM @ FP16

156 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

78B model exceeds a single RTX 4090's 24GB VRAM even at Q4 (~53GB needed); no single-4090 measurement exists.

Quantizations

AWQ

InternVL 2.5 family

Elo progression across releases

InternVL2.5-78BDec 20241350.1

Resources

HFOpenGVLab/InternVL2_5-78B GHOpenGVLab/InternVL APIAvailable via OpenGVLab (Shanghai AI Lab) API

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