UForm Model Benchmarks

October 30, 2025 · View on GitHub

Accuracy

Embedding Models

Few retrieval benchmarks exist for multimodal embeddings. The most famous ones for English are "MS-COCO" and "Flickr30k". Evaluating uform-vl-english model, one can expect the following numbers for search quality.

DatasetRecall @ 1Recall @ 5Recall @ 10
Flickr0.7270.9150.949
MS-COCO ¹0.5100.7610.838

For multilingual benchmarks, we've created the unum-cloud/coco-sm repository². Evaluating the unum-cloud/uform-vl-multilingual-v2 model, one can expect the following metrics for text-to-image search, compared against xlm-roberta-base-ViT-B-32 OpenCLIP model.

LanguageOpenCLIP @ 1UForm @ 1OpenCLIP @ 5UForm @ 5OpenCLIP @ 10UForm @ 10Speakers
English 🇺🇸37.837.763.565.073.575.91'452 M
Chinese 🇨🇳27.332.251.359.062.170.51'118 M
Hindi 🇮🇳20.731.342.557.953.769.6602 M
Spanish 🇪🇸32.635.658.062.868.873.7548 M
Arabic 🇸🇦22.731.744.957.855.869.2274 M
French 🇫🇷31.335.456.562.667.473.3274 M

All languages:

LanguageOpenCLIP @ 1UForm @ 1OpenCLIP @ 5UForm @ 5OpenCLIP @ 10UForm @ 10Speakers
Arabic 🇸🇦22.731.744.957.855.869.2274 M
Armenian 🇦🇲5.622.014.344.720.256.04 M
Chinese 🇨🇳27.332.251.359.062.170.51'118 M
English 🇺🇸37.837.763.565.073.575.91'452 M
French 🇫🇷31.335.456.562.667.473.3274 M
German 🇩🇪31.735.156.962.267.473.3134 M
Hebrew 🇮🇱23.726.746.351.857.063.59 M
Hindi 🇮🇳20.731.342.557.953.769.6602 M
Indonesian 🇮🇩26.930.751.457.062.768.6199 M
Italian 🇮🇹31.334.956.762.167.173.167 M
Japanese 🇯🇵27.432.651.559.262.670.6125 M
Korean 🇰🇷24.431.548.157.859.269.281 M
Persian 🇮🇷24.028.847.054.657.866.277 M
Polish 🇵🇱29.233.653.960.164.771.341 M
Portuguese 🇵🇹31.632.757.159.667.971.0257 M
Russian 🇷🇺29.933.954.860.965.872.0258 M
Spanish 🇪🇸32.635.658.062.868.873.7548 M
Thai 🇹🇭21.528.743.054.653.766.061 M
Turkish 🇹🇷25.533.049.159.660.370.888 M
Ukranian 🇺🇦26.030.649.956.760.968.141 M
Vietnamese 🇻🇳25.428.349.253.960.365.585 M
Mean26.5±6.431.8±3.549.8±9.858.1±4.560.4±10.669.4±4.3-
Google Translate27.4±6.331.5±3.551.1±9.557.8±4.461.7±10.369.1±4.3-
Microsoft Translator27.2±6.431.4±3.650.8±9.857.7±4.761.4±10.668.9±4.6-
Meta NLLB24.9±6.732.4±3.547.5±10.358.9±4.558.2±11.270.2±4.3-

Generative Models

ModelLLM SizeSQAMMEMMBenchAverage¹
UForm-Gen2-Qwen-500m0.5B45.5880.142.029.31
MobileVLM v21.4B52.11302.857.736.81
LLaVA-Phi2.7B68.41335.159.842.95

For captioning evaluation we measure CLIPScore and RefCLIPScore³.

ModelSizeCaption LengthCLIPScoreRefCLIPScore
llava-hf/llava-1.5-7b-hf7BLong0.8780.529
llava-hf/llava-1.5-7b-hf7BShort0.8860.531
Salesforce/instructblip-vicuna-7b7BLong0.9020.534
Salesforce/instructblip-vicuna-7b7BShort0.8480.523
unum-cloud/uform-gen1.5BLong0.8470.523
unum-cloud/uform-gen1.5BShort0.8420.522
unum-cloud/uform-gen-chat1.5BLong0.8600.525
unum-cloud/uform-gen-chat1.5BShort0.8580.525

Results for VQAv2 evaluation.

ModelSizeAccuracy
llava-hf/llava-1.5-7b-hf7B78.5
unum-cloud/uform-gen1.5B66.5

¹ Train split was in training data.
² Lacking a broad enough evaluation dataset, we translated the COCO Karpathy test split with multiple public and proprietary translation services, averaging the scores across all sets, and breaking them down in the bottom section.
³ We used apple/DFN5B-CLIP-ViT-H-14-378 CLIP model.

Speed

Embedding Models

UForm comes pre-packaged with speed benchmarks for the models.

$ python python/scripts/bench_encoders.py --help
usage: bench_encoders.py [-h] [--filter-out FILTER_OUT] [--batch-size BATCH_SIZE]

options:
  -h, --help            show this help message and exit
  --filter-out FILTER_OUT
                        Filter out models, backends, or devices with a Regular Expression.
  --batch-size BATCH_SIZE
                        Batch size for the benchmark. Batch size 1 measures latency. Large batch sizes may not fit on every GPU.

On Nvidia B200, the results may look like:

uv run python python/scripts/bench_encoders.py --batch-size 2048 --gpu --torch
ModelDeviceBackendPrecisionImages/sTexts/s
unum-cloud/uform3-image-text-english-basecudatorchbfloat166662.738482.7
unum-cloud/uform3-image-text-english-largecudatorchbfloat162930.253927.5
unum-cloud/uform3-image-text-english-smallcudatorchbfloat161385.66611.2
unum-cloud/uform3-image-text-multilingual-basecudatorchbfloat167235.136690.4

On the 160-core dual-socket Intel Emerald Rapids CPU-only setup, the results may look like:

uv run python python/scripts/bench_encoders.py --batch-size 128 --cpu --torch --onnx
ModelDeviceBackendPrecisionImages/sTexts/s
unum-cloud/uform3-image-text-english-basecputorchbfloat16164.33228.1
unum-cloud/uform3-image-text-english-basecpuonnxunknown212.81752.8
unum-cloud/uform3-image-text-english-largecputorchbfloat1632.13550.8
unum-cloud/uform3-image-text-english-largecpuonnxunknown58.91348.2
unum-cloud/uform3-image-text-english-smallcputorchbfloat16335.95801.1
unum-cloud/uform3-image-text-english-smallcpuonnxunknown325.42589.3
unum-cloud/uform3-image-text-multilingual-basecputorchbfloat16153.24026.2
unum-cloud/uform3-image-text-multilingual-basecpuonnxunknown197.51401.4

Generative Models

$ python python/scripts/bench_decoders.py --help
usage: bench_decoders.py [-h] [--filter-out FILTER_OUT] [--batch-size BATCH_SIZE]

options:
  -h, --help            show this help message and exit
  --batch-size BATCH_SIZE
                        Batch size for the benchmark. Batch size 1 measures latency. Large batch sizes may not fit on every GPU.
  --max-length MAX_LENGTH
                        Maximum length of the generated text in tokens.

On Nvidia H100 GPU, the following performance is expected on text token generation using float16, equivalent PyTorch settings, and greedy decoding.

ModelSizeDecoding SpeedDecoding Parallel Streams
llava-hf/llava-1.5-7b-hf7 B~ 141 tokens/s~ 4 K tokens/s (32 streams)
Salesforce/instructblip-vicuna-7b7 B~ 211 tokens/s~ 2 K tokens/s (32 streams)
unum-cloud/uform-gen1.5 B~ 252 tokens/s~ 3 K tokens/s (128 streams)
unum-cloud/uform-gen2-dpo1.2 B~ 372 tokens/s~ 10 K tokens/s (64 streams)

On Nvidia RTX 3090, the following performance is expected on text token generation using float16, equivalent PyTorch settings, and greedy decoding.

ModelSizeDecoding SpeedSpeedup
llava-hf/llava-1.5-7b-hf7 B~ 40 tokens/s
Salesforce/instructblip-vicuna-7b7 B~ 40 tokens/s
unum-cloud/uform-gen1.5 B~ 140 tokens/sx 3.5