Normalized evidence matrix

MiniMax H3 Benchmark Evidence

These records are normalized, not ranked. Runtime rows are comparable only when GPU count, backend, mode, resolution, duration, steps, acceleration and environment materially match.

Evidence checked August 11, 2026

10 evidence records

GPU / gradeEnvironmentWorkloadTotal runtimeSystem RAMEvidence
RTX 3060
A_claim
ComfyUI
— · —
duration —
Source
RTX 3060
B_reported
ComfyUI
Windows 10
R2V · 1344x768
5s
10 min32 GBSource
generic consumer GPU
A_fact
Diffusers
— · 960x544
duration —
≈75 GBSource
2× RTX 5090
A_benchmark
SGLang
T2VA · 1344x768
5s
9.3 min377 GiBSource
2× RTX 5090
A_benchmark
SGLang
T2VA · 1344x768
5s
1.3 minSource
RTX 3090
B_reproduced
ComfyUI
Linux
T2V · 832x480
15.083333s
23.3 min31 GBSource
RTX 3090
B_reported
ComfyUI
Ubuntu Server 26.04
— · 0.4 MP
15s
64 GBSource
RX 9070 XT
B_reported
ComfyUI ROCm
Linux
T2V · 848x480
5.2s
6.7 min64 GBSource
RX 9070 XT
B_reported
ComfyUI ROCm
Linux
T2V · 1344x768
5.2s
22.3 min64 GBSource
RTX 5070 Ti
B_reported
ComfyUI
Windows 11 Pro 24H2
I2V · 0.4 MP
5s
48 GBSource

“—” means not recorded. It is not zero. Acceleration variants and workarounds remain attached to the source record in the underlying data pack.