Benchmark Explorer

Model FamilyTPU GenerationTask

			TPU	Weight dtype	I/O dtype
Cosmos3 Edge (4B)	T2V	832x480	TPU v4	bf16	bf16	2.909	274.8	0.440	17.21	5	FlyingGiraffe
CogVideoX 2B	T2V	720x480	TPU v4	bf16	bf16	4.230	519.2	0.094	17.18	5	FlyingGiraffe
LTX-2.5 22B (distilled)	T2V	1216x704	TPU v4	bf16	bf16	4.663	215.0	0.563	15.35	5	FlyingGiraffe
LTX-Video 13B (dev)	T2V	1216x704	TPU v4	bf16	bf16	5.200	347.6	0.348	15.27	5	FlyingGiraffe
LTX-Video 2B (distilled)	T2V	1216x704	TPU v4	bf16	bf16	5.916	162.6	0.744	8.84	5	FlyingGiraffe
Wan2.1 1.3B	T2V	832x480	TPU v4	bf16	bf16	6.965	456.4	0.177	10.19	5	FlyingGiraffe
Cosmos3 Nano (16B)	T2V	1280x704	TPU v4	bf16	bf16	7.141	437.5	0.213	29.49	5	FlyingGiraffe
LTX-2.5 22B (dev)	T2V	1216x704	TPU v4	bf16	bf16	7.262	409.7	0.295	16.69	5	FlyingGiraffe
CogVideoX 5B	I2V	720x480	TPU v4	bf16	bf16	9.413	851.1	0.058	23.33	5	FlyingGiraffe
CogVideoX 5B	T2V	720x480	TPU v4	bf16	bf16	9.413	835.3	0.059	23.18	5	FlyingGiraffe
Wan2.2 TI2V-5B	T2V	1280x704	TPU v4	fp32	bf16	10.518	649.7	0.186	18.29	5	FlyingGiraffe
Wan2.2 TI2V-5B	I2V	800x1088	TPU v4	fp32	bf16	12.068	665.4	0.182	18.26	5	FlyingGiraffe
LTX-Video 13B (distilled)	T2V	1216x704	TPU v4	bf16	bf16	13.028	300.6	0.403	15.27	5	FlyingGiraffe
Wan2.1 14B	T2V	832x480	TPU v4	bf16	bf16	26.136	1552.9	0.052	17.17	5	FlyingGiraffe
Wan2.1 14B	I2V	544x720	TPU v4	bf16	bf16	28.132	1372.1	0.059	22.08	5	FlyingGiraffe
Cosmos-Predict2.5 2B	T2V	1280x704	TPU v4	bf16	bf16	38.779	1497.3	0.062	16.03	5	FlyingGiraffe
Wan2.2 A14B	T2V	832x480	TPU v4	fp32	bf16	43.182	2408.2	0.034	28.38	5	FlyingGiraffe
Wan2.2 A14B	I2V	544x720	TPU v4	fp32	bf16	44.499	2063.2	0.039	28.35	5	FlyingGiraffe
Wan2.2 A14B	I2V	832x1104	TPU v4	fp32	bf16	49.062	2163.7	0.015	20.52	5	FlyingGiraffe
CogVideoX1.5 5B	I2V	1360x768	TPU v4	bf16	bf16	52.796	3400.8	0.024	31.45	5	FlyingGiraffe
CogVideoX1.5 5B	T2V	1360x768	TPU v4	bf16	bf16	52.796	3380.2	0.024	31.46	5	FlyingGiraffe
LTX-2.5 22B (dev, diffusion VAE)	T2V	1216x704	TPU v4	bf16	bf16	95.320	3444.3	0.035	16.11	5	FlyingGiraffe
HunyuanVideo-1.5 8.3B (480p)	I2V	544x720	TPU v4	bf16	bf16	112.880	3838.1	0.032	31.44	5	FlyingGiraffe
HunyuanVideo-1.5 8.3B (480p)	T2V	832x480	TPU v4	bf16	bf16	119.459	4015.2	0.030	29.14	5	FlyingGiraffe
Wan2.1 14B	T2V	1280x720	TPU v4	fp32	bf16	123.011	6562.9	0.012	23.00	5	FlyingGiraffe
Wan2.1 14B	I2V	832x1104	TPU v4	fp32	bf16	127.249	5651.6	0.014	32.70	5	FlyingGiraffe
Cosmos-Predict2.5 14B	T2V	1280x704	TPU v4	bf16	bf16	127.988	4577.1	0.020	14.70	5	FlyingGiraffe
HunyuanVideo-1.5 8.3B (720p)	I2V	832x1104	TPU v4	bf16	bf16	219.186	7125.0	0.017	32.00	5	FlyingGiraffe
HunyuanVideo-1.5 8.3B (720p)	T2V	1280x720	TPU v4	bf16	bf16	220.993	7133.5	0.017	30.28	5	FlyingGiraffe
HunyuanVideo 13B	T2V	1280x720	TPU v4	bf16	bf16	299.614	15620.7	0.008	18.40	5	FlyingGiraffe
HunyuanVideo 13B	I2V	832x1088	TPU v4	bf16	bf16	306.369	16007.4	0.008	18.41	5	FlyingGiraffe
LTX-2.5 22B (distilled, diffusion VAE)	T2V	1216x704	TPU v4	bf16	bf16	335.095	3247.4	0.037	14.76	5	FlyingGiraffe

Resolution is always width×height. T2V rows are landscape (width > height); I2V rows are portrait (height > width) — output resolution is derived from the standardized conditioning image's own aspect ratio, not a fixed size. The exception is CogVideoX-5b-I2V and CogVideoX1.5-5B-I2V, which are locked to one fixed generation resolution (same as their T2V sibling) and rescale the rendered video to the conditioning image's aspect ratio afterward — their rows show that fixed generation resolution, not the final rescaled output size.

Submit Your Own Results

Ran vidax on hardware not covered above, or want to add a newer model/config? Submit the following form to add your results. Every row shows its contributor linked to their GitHub profile.

Submit a benchmark result on GitHub

The form covers one result at a time. Submitting several at once (or you already have them in this shape)? Skip the individual fields and paste a JSON array into the form's Bulk JSON field instead — each object:

[
  {
    "family": "Wan2.1",
    "version": "2.1",
    "size": "1.3b",
    "sizeLabel": "1.3B",
    "task": "T2V",
    "resolution": "832x480",
    "numFrames": 81,
    "numSteps": 50,
    "jaxVersion": "0.11.0",
    "deviceKind": "TPU v7",
    "deviceTopology": "2x2x4",
    "tensorParallelSize": 2,
    "sequenceParallelSize": 1,
    "ioDtype": "bf16",
    "weightDtype": "bf16",
    "numRuns": 1,
    "compileS": 62.1,
    "perStepS": 2.41,
    "wallS": 168.3,
    "peakHbmGb": 9.8
  }
]