Changes
What changed on model pages, day by day: new models and sizes, rows on other runtimes, first scores, and numbers we measured again. Edits to text aren't listed.
Days are in UTC. A number that changed is ours, measured again by the same protocol, and names where it ran: the RTX 4090 on PyTorch or TensorRT, or the CPU with no GPU. How we verify
CPU-only rows for 44 models
Show all 44 models
- BEN2
- CPU-only rows (PyTorch and ONNX Runtime): base; no number for base on ONNX Runtime
- BiRefNet
- CPU-only rows (PyTorch): general, hr, and lite
- D-FINE
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- DEIMv2
- CPU-only rows (PyTorch): s, m, l, and x
- Depth Anything 3 Metric
- CPU-only row (PyTorch): large, without a number
- Depth Anything 3 Mono
- CPU-only row (PyTorch): large, without a number
- Depth Anything V2
- CPU-only rows (PyTorch): small, base, and large
- EoMT
- CPU-only rows (PyTorch): large-640 and large-1280
- FLUX.2 [klein]
- CPU-only rows (PyTorch): 4b and base-4b; no number for base-4b on PyTorch
- Gemma 4
- CPU-only rows (llama.cpp): 12b-q4_0 and 31b-q4_0
- Grounding DINO
- CPU-only rows (PyTorch): tiny and base
- GTR
- CPU-only rows (PyTorch): s, m, l, and x, without a number
- GTR-Pose
- CPU-only rows (PyTorch): s, m, l, and x, without a number
- GTR-Seg
- CPU-only rows (PyTorch): s, m, l, and x, without a number
- LLMDet
- CPU-only rows (PyTorch): tiny, base, and large
- LT-DETR
- CPU-only rows (PyTorch and ONNX Runtime): s, m, and l
- Lumina-Image 2.0
- CPU-only row (PyTorch): 2.0, without a number
- MM-Grounding-DINO
- CPU-only rows (PyTorch): tiny, base, and large
- MoGe-2
- CPU-only rows (PyTorch): vits, vitb, and vitl
- PixArt-Σ
- CPU-only row (PyTorch): xl-2-1024-ms, without a number
- Playground v2.5
- CPU-only row (PyTorch): 1024px-aesthetic, without a number
- Qwen3
- CPU-only rows (llama.cpp): 0.6b-q8_0 and 8b-q4_k_m
- Qwen3.5
- CPU-only row (llama.cpp): 9b-q4_k_m
- Qwen3.8
- CPU-only row (llama.cpp): 27b-q4_k_m
- RF-DETR
- CPU-only rows (PyTorch): nano, small, medium, and large
- RF-DETR Keypoint
- CPU-only row (PyTorch): preview
- RF-DETR-Seg
- CPU-only rows (PyTorch): nano, small, medium, large, xlarge, and 2xlarge
- RT-DETRv4
- CPU-only rows (PyTorch and ONNX Runtime): s, m, l, and x
- Sana 1.5
- CPU-only rows (PyTorch): 1.6b and 4.8b, without a number
- Sana-Sprint
- CPU-only row (PyTorch): 1.6b
- SDXL-Lightning
- CPU-only row (PyTorch): 4step
- SDXL-Turbo
- CPU-only row (PyTorch): 1.0
- Stable Diffusion 1.5
- CPU-only row (PyTorch): v1-5
- Stable Diffusion XL
- CPU-only row (PyTorch): base, without a number
- YOLO11
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLO11-pose
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLO11-seg
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLO26
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLO26-pose
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLO26-seg
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLOv8
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLOv8-pose
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- YOLOv8-seg
- CPU-only rows (PyTorch and ONNX Runtime): n, s, m, l, and x
- Z-Image
- CPU-only rows (PyTorch): turbo and base, without a number
- BEN2
First scores for FLUX.2 [klein]
- FLUX.2 [klein]
- First GenEval score: base-4b
- FLUX.2 [klein]
8 new models; new sizes of FLUX.2 [klein] and Z-Image
- Lumina-Image 2.0
- Added
- PixArt-Σ
- Added
- Playground v2.5
- Added
- Sana 1.5
- Added, with sizes 1.6b and 4.8b
- Sana-Sprint
- Added
- SDXL-Lightning
- Added
- SDXL-Turbo
- Added
- Stable Diffusion 1.5
- Added
- FLUX.2 [klein]
- New size: base-4b
- Z-Image
- New size: base
- Lumina-Image 2.0
TensorRT rows for D-FINE; new numbers for D-FINE
- D-FINE
- TensorRT 10.13.2.6 rows: s, m, and l
- n re-measured · RTX 4090: 10.4 ms (TensorRT 10.16.1.11 fp16, 640x640) → 10.5 ms (TensorRT 10.13.2.6 fp16, 640x640)
- n re-scored · box AP (TensorRT): 42.5 → 42.6
- D-FINE
4 new models; TensorRT rows for D-FINE and YOLO26
- FLUX.2 [klein]
- Added
- LT-DETR
- Added, with sizes s, m, and l
- Stable Diffusion XL
- Added
- Z-Image
- Added
- D-FINE
- TensorRT 10.16.1.11 row: n
- YOLO26
- TensorRT 10.16.1.11 rows: s, m, l, and x
- FLUX.2 [klein]
11 new models; new sizes of Depth Anything V2; new numbers for Qwen3 and RF-DETR
Show all 14 models
- Depth Anything 3 Metric
- Added
- Depth Anything 3 Mono
- Added
- Gemma 4
- Added, with sizes 12b-q4_0 and 31b-q4_0
- Qwen3.5
- Added
- Qwen3.8
- Added
- YOLO11
- Added, with sizes n, s, m, l, and x
- YOLO11-pose
- Added, with sizes n, s, m, l, and x
- YOLO11-seg
- Added, with sizes n, s, m, l, and x
- YOLOv8
- Added, with sizes n, s, m, l, and x
- YOLOv8-pose
- Added, with sizes n, s, m, l, and x
- YOLOv8-seg
- Added, with sizes n, s, m, l, and x
- Depth Anything V2
- New sizes: base and large
- Qwen3
- 0.6b-q8_0 re-scored · GSM8K: 43.1 → 43.3
- 8b-q4_k_m re-scored · GSM8K: 86.3 → 89.6
- RF-DETR
- nano re-measured · RTX 4090 · PyTorch fp16, 384x384: 9.9 → 9.8 ms
- Depth Anything 3 Metric
Site launched with 21 models
The public site went up with 21 models, each run and measured on our RTX 4090. See every model