List your company

Multimodal Evaluation on MME (MME-P and MME-C)

1,864MME-P ScoreTwigVLM
1,323.36641,463.72321,604.081,744.4368Mar 18, 2025May 27, 2025Aug 5, 2025Oct 14, 2025Dec 23, 2025Mar 3, 2026May 13, 2026

Evaluation results

MethodLinks
TwigVLMRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,864—
TwigVLM++Retained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,858—
LLaVA-NeXT-7BRetained Tokens=2880, Base Model=LLaVA-NeXT-7B2025.031,851—
FastVRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,807—
VisionZipRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,787—
VisionZip‡Retained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,778—
TwigVLM++Retained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,773—
SparseVLMRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.031,772—
VisionZip‡Retained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,770—
TwigVLMRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,758—
VisionZipRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,702—
FP16 (Baseline)Backbone=Qwen 2.5 VL 7B, Avg. Bit width=162025.091,695640
SparseVLMRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,694—
OmniQuantBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.091,655610
Qwen2.5VL 7B# Vis tok.=14002025.041,652.7633.4
AWQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.091,645620
MustDropRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,641—
LUQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=2.75, Quantized Layers=12 layer2025.091,640600
FLARE-X 8B# Vis tok.=14002025.041,639.1378.9
GPTQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.091,638610
ETAModel=InternLM-XComposer-2.52025.101,629.4546.1
LLaVA-OneVision 7B# Vis tok.=24002025.041,626483
VanillaModel=InternLM-XComposer-2.52025.101,623.7551.1
ECSOModel=InternLM-XComposer-2.52025.101,623.7551.1
MoRASModel=InternLM-XComposer-2.52025.101,623.7551.1
ASTRAModel=InternLM-XComposer-2.52025.101,617.4546.8
FLARE-X 3B# Vis tok.=14002025.041,612.4437.2
CoCAModel=InternLM-XComposer-2.52025.101,606.5551.1
MiniCPM-Llama3-V-2.5 8B# Vis tok.=4002025.041,604.8398.5
FLARE-L 8B# Vis tok.=6302025.041,596.8347.3
Qwen2.5VL 3B# Vis tok.=14002025.041,592.4607.5
AutoSteerModel=InternLM-XComposer-2.52025.101,591.2544.3
FLARE-L 3B# Vis tok.=6302025.041,588.3425.6
FigStepModel=InternLM-XComposer-2.52025.101,577.7516.8
MoRASModel=LLaVA-OneVision-7B2025.101,561.8409.6
VanillaModel=LLaVA-OneVision-7B2025.101,560.9409.6
ECSOModel=LLaVA-OneVision-7B2025.101,560.9409.6
ETAModel=LLaVA-OneVision-7B2025.101,560.9409.6
Florence-VL 8B# Vis tok.=5762025.041,560381.1
Eagle 8B# Vis tok.=10242025.041,559—
DeepSeek-VL2-Tiny# Vis tok.=15002025.041,548.3357.1
AutoSteerModel=LLaVA-OneVision-7B2025.101,548.1409.6
Cambrain 8B# Vis tok.=5762025.041,547.1—
FastVRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.031,539—
Molmo 7B# Vis tok.=12002025.041,536.1341.6
MiniCPM-V-2.0 3B# Vis tok.=4002025.041,531.4322.8
ETABackbone=LLaVA-1.5-13B2025.101,531.2296.1
VanillaBackbone=LLaVA-1.5-13B2025.101,529.9298.6
ECSOBackbone=LLaVA-1.5-13B2025.101,529.9298.6
MoRASBackbone=LLaVA-1.5-13B2025.101,529.9298.6
CoCAModel=LLaVA-1.5-7B2025.101,526.5283.6
FullData %=100%2026.051,517.3656.1
GRACEData %=20%2026.051,512.3682.9
LESSData %=20%2026.051,511.4668.6
AutoSteerBackbone=LLaVA-1.5-13B2025.101,510.9296.4
FP16 (Baseline)Backbone=LLaVA-1.5 7B, Avg. Bit width=162025.091,510350
ETAModel=LLaVA-1.5-7B2025.101,509.3339.6
GRACEData %=5%2026.051,509.3617.5
LLaVA1.5-7BPruning budget (keep tokens)=Full (No pruning), Base model=LLaVA1.5-7B2026.041,509.13—
VanillaModel=LLaVA-1.5-7B2025.101,505.1355.7
ECSOModel=LLaVA-1.5-7B2025.101,505.1355.7
MoRASModel=LLaVA-1.5-7B2025.101,505.1355.7
GRACEData %=10%2026.051,500.7660.4
GRACEData %=15%2026.051,500.5679.3
CADCData %=20%2026.051,498.9630.7
Florence-VL 3B# Vis tok.=5762025.041,498.7403.9
ICONSData %=20%2026.051,497675
RandomData %=20%2026.051,495.5653.2
LongestData %=20%2026.051,494.5687.5
VanillaModel=Qwen-VL-Chat2025.101,489.9331.8
ECSOModel=Qwen-VL-Chat2025.101,489.9331.8
MoRASModel=Qwen-VL-Chat2025.101,489.9331.8
ETAModel=Qwen-VL-Chat2025.101,487.9331.8
FigStepModel=Qwen-VL-Chat2025.101,480.9296.4
AutoSteerModel=LLaVA-1.5-7B2025.101,479.9315.4
StepmaxData %=20%2026.051,477.7637.5
AutoSteerModel=Qwen-VL-Chat2025.101,474.5348.2
Phi 3.5-Vision# Vis tok.=21002025.041,473.4412.1
CoCABackbone=LLaVA-1.5-13B2025.101,472.8301.8
FigStepModel=LLaVA-OneVision-7B2025.101,472.2395.4
ASTRAModel=LLaVA-1.5-7B2025.101,472331.4
OmniQuantBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.091,466318
ASTRAModel=Qwen-VL-Chat2025.101,465.9323.2
SlimLLMBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.091,465345
CoCAModel=LLaVA-OneVision-7B2025.101,463.4403.2
AWQBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.091,456349
GPTQBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.091,450347
FigStepModel=LLaVA-1.5-7B2025.101,435.7275
SlimLLMBackbone=LLaVA-1.5 7B, Avg. Bit width=32025.091,429329
MI-AttentionPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.041,428.03—
FigStepBackbone=LLaVA-1.5-13B2025.101,423.9322.1
ASTRAModel=LLaVA-OneVision-7B2025.101,406.4387.9
MI-PrunerPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.041,381.35—
CoCAModel=Qwen-VL-Chat2025.101,377.1319.3
ASTRABackbone=LLaVA-1.5-13B2025.101,371.8327.5
VisPrunerPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.041,369.9—
VisionZipPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.041,365.6—
LUQBackbone=LLaVA-1.5 7B, Avg. Bit width=2.54, Quantized Layers=16 layer2025.091,365257
GPTQBackbone=LLaVA-1.5 7B, Avg. Bit width=32025.091,346273
MI-AttentionPruning budget (keep tokens)=32, Base model=LLaVA1.5-7B2026.041,344.16—