openbmb/MiniCPM-o-4_5
Any-to-Any
ā¢
9B
ā¢
Updated
ā¢
206
ā¢
172
Hebangwen
hebangwen
AI & ML interests
None yet
Recent Activity
liked
a model
about 5 hours ago
openbmb/MiniCPM-o-4_5
liked
a model
1 day ago
openbmb/AgentCPM-Report
liked
a model
1 day ago
stepfun-ai/Step-3.5-Flash
Organizations
liked
a
model
about 5 hours ago
liked
2
models
1 day ago
upvoted
a
paper
5 days ago
llama.cpp support please.
ā¤ļø
12
2
#1 opened 7 days ago
by
rosspanda0
liked
a
model
7 days ago
Some question about Inference Benchmark
1
#8 opened 9 days ago
by
hebangwen
liked
a
model
9 days ago
liked
a
model
22 days ago
liked
a
model
24 days ago
reacted to
Reality123b's
post with š¤
26 days ago
Post
2104
Happy birthday to me!!!
- 2 replies
reacted to
yeonseok-zeticai's
post with š
about 1 month ago
Post
341
ā” MobileCLIP2 Complete On-device Study: TMLR 2025 Featured Model on Mobile
Major Release: Comprehensive mobile deployment study of Apple's MobileCLIP2 (TMLR August 2025 Featured) with detailed performance benchmarks across 52+ mobile devices!
šÆ Model Overview:
- Architecture: Multi-modal reinforced training (vision + language)
- Research: TMLR 2025 Featured Certification
- Innovation: Improved efficiency-accuracy trade-offs vs SigLIP/OpenAI CLIP
- Specialty: Zero-shot image classification and retrieval
š Mobile Performance Results:
Latency Metrics:
- NPU (Best): 9.74ms average inference
- GPU: 39.00ms average
- CPU: 494.89ms average
- NPU Advantage: 115.94x speedup over CPU baseline!
Memory Efficiency:
- Model Size: 1.66 GB (production optimized)
- Runtime Memory: 466.18 MB peak consumption
- Load Range: 0-1,884 MB across device categories
- Inference Range: 431-1,616 MB
Accuracy Preservation:
- FP16 Precision: 39.78 dB maintained
- Quantized Mode: 15.07 dB (INT optimization available)
- Zero-shot Quality: Production-grade vision-language matching
š Research Highlights:
MobileCLIP2-S4 Performance:
- Matches SigLIP-SO400M/14 accuracy
- 2x fewer parameters
- 2.5x lower latency than DFN ViT-L/14
MobileCLIP-S0 Efficiency:
- Similar zero-shot performance to OpenAI ViT-B/16
- 4.8x faster inference
- 2.8x smaller model size
MobileCLIP-S2 Advantages:
- Better avg zero-shot than SigLIP ViT-B/16
- 2.3x faster, 2.1x smaller
- Trained with 3x less seen samples
MobileCLIP-B (LT) Accuracy:
- 77.2% ImageNet zero-shot
- Surpasses OpenAI ViT-L/14@336
- Better than DFN and SigLIP similar architectures
š Resources:
- Complete Study: https://mlange.zetic.ai/p/Steve/MobileCLIP2-image
Ready to build vision-language applications that run entirely on-device?
The future of multi-modal AI runs locally in everyone's pocket! š
Major Release: Comprehensive mobile deployment study of Apple's MobileCLIP2 (TMLR August 2025 Featured) with detailed performance benchmarks across 52+ mobile devices!
šÆ Model Overview:
- Architecture: Multi-modal reinforced training (vision + language)
- Research: TMLR 2025 Featured Certification
- Innovation: Improved efficiency-accuracy trade-offs vs SigLIP/OpenAI CLIP
- Specialty: Zero-shot image classification and retrieval
š Mobile Performance Results:
Latency Metrics:
- NPU (Best): 9.74ms average inference
- GPU: 39.00ms average
- CPU: 494.89ms average
- NPU Advantage: 115.94x speedup over CPU baseline!
Memory Efficiency:
- Model Size: 1.66 GB (production optimized)
- Runtime Memory: 466.18 MB peak consumption
- Load Range: 0-1,884 MB across device categories
- Inference Range: 431-1,616 MB
Accuracy Preservation:
- FP16 Precision: 39.78 dB maintained
- Quantized Mode: 15.07 dB (INT optimization available)
- Zero-shot Quality: Production-grade vision-language matching
š Research Highlights:
MobileCLIP2-S4 Performance:
- Matches SigLIP-SO400M/14 accuracy
- 2x fewer parameters
- 2.5x lower latency than DFN ViT-L/14
MobileCLIP-S0 Efficiency:
- Similar zero-shot performance to OpenAI ViT-B/16
- 4.8x faster inference
- 2.8x smaller model size
MobileCLIP-S2 Advantages:
- Better avg zero-shot than SigLIP ViT-B/16
- 2.3x faster, 2.1x smaller
- Trained with 3x less seen samples
MobileCLIP-B (LT) Accuracy:
- 77.2% ImageNet zero-shot
- Surpasses OpenAI ViT-L/14@336
- Better than DFN and SigLIP similar architectures
š Resources:
- Complete Study: https://mlange.zetic.ai/p/Steve/MobileCLIP2-image
Ready to build vision-language applications that run entirely on-device?
The future of multi-modal AI runs locally in everyone's pocket! š
liked
a
model
about 1 month ago
liked
a
model
about 2 months ago
How can we access the acoustic encoder and semantics encoder?
1
#20 opened about 2 months ago
by
hebangwen
liked
2
models
about 2 months ago
Humble request for a stable vLLM/SGLang deployment setup for DeepSeek-V3.2
š
ā¤ļø
13
11
#15 opened 2 months ago
by
burrowswang
make deepseekocr compatible with mps and latest transformers
#96 opened 2 months ago
by
hebangwen
