Model

CogVLM-17B

OPEN

Zhipu AI / THUDM (Tsinghua) · CogVLM family · released Oct 1, 2023

Open VLM (10B vision expert + 7B Vicuna-based LM) with SOTA-level results on 10 cross-modal benchmarks at 490x490 resolution.

ReasoningCodingVisionFunction callingTool useAgentic

426.2

Elo · unrated

Parameters

17B

Active params

17B (dense)

Context

4K tokens

Architecture

Dense dual-tower vision-language Transformer (trainable visual expert modules)

License

Apache-2.0 (code); CogVLM Model License for weights (non-commercial use / commercial by agreement)

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

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

COCO CaptionsVision148.7#1

best: this model · 148.7

Flickr30kVision94.9#1

best: this model · 94.9

MathVistaVision34.5%#60

best: Seed 2.1 Pro · 90.7%

MMBenchVision77.6%#3

best: Phi-3.5-vision (4.2B) · 81.9%

MMMUVision41.1%#104

best: Claude Fable 5 · 89.3%

OK-VQAVision64.8%#1

best: this model · 64.8%

SEED-BenchVision72.5%#4

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

TextVQAVision70.4%#22

best: Molmo 2 8B · 85.7%

VQAv2Vision82.3%#4

best: Reka Edge (2026) · 88.4%

Run it locally

VRAM @ Q4

11 GB

VRAM @ FP16

—

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

Throughput data unavailable.

Quantizations

INT4 · INT8

Fine-tune it

Research only

QLoRA12.7 GB1× RTX 4070 Ti 16GB

LoRA37.4 GB1× A100 80GB

Full fine-tune274.0 GB2× H200 141GB

QLoRA SFT on ~10k samples ≈ $6.22 (1× RTX 4070 Ti 16GB)

Plan a fine-tune →

CogVLM family

Elo progression across releases

CogVLM-17BOct 2023426.2

Resources

HFTHUDM/cogvlm-chat-hf GHzai-org/CogVLM

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