The LMSYS team published a MiniMax-H3 inference performance report on August 27, testing on 8x NVIDIA H200 GPUs with 141GB of memory each, running the SGLang Diffusion framework. Parameters were fixed at 1344×768 resolution, 24 FPS, and 50 denoising steps, producing both 5-second and 10-second clips, benchmarked against Diffusers.
The findings come in two tiers. The first one holds nothing back:
"SGLang's dense, lossless path is 1.85–1.95× faster than Diffusers with no approximation: the same denoising work, on a faster runtime."
On the text-to-video animation pipeline, the lossless tier runs 1.87 to 1.95x faster than Diffusers; on the frame-list-to-video pipeline, it's 1.91 to 1.95x faster. Both pipelines come close to doubling in speed, and SSIM sits at an exact 1.0000 — pixel-identical output.
Three layers of speedup stack up to 6x
The report breaks the speedup into three components, each with a different character.
Fused operators are pure engineering. SGLang merges the AdaLN update, gated residual, SwiGLU activation, and QK RMSNorm with 3D RoPE into a single kernel, cutting down on memory round-trips. Isolated operator-level tests show speedups between 2.00x and 12.16x; end-to-end, that translates into the 1.95x figure above. This layer changes none of the computed results.
Cache-DiT is where approximation begins: adjacent denoising steps reuse cached results, with a residual-difference threshold deciding whether to reuse. The conservative threshold is 0.04, the more aggressive "stride" tier is 0.08. The conservative tier reaches 2.65x and 2.99x across the two pipelines, with SSIM dropping to around 0.90 and 0.94.
SubBlock sparse attention is the most aggressive: a block-sparse router trims which KV blocks participate in computation after step 10, with sparsity set to 0.75 or 0.80. Stacked with the Cache-DiT stride tier, text-to-video reaches 4.90x to 5.72x, and frame-list-to-video reaches 5.64x to 6.24x.
What the 6x tier costs in image quality
The cost shows up entirely in SSIM, and the two pipelines diverge sharply. Frame-list-to-video holds up well: at 0.75 sparsity plus stride, the 10-second clip's SSIM is still 0.9202, and the 5-second clip's is 0.8629. Text-to-video drops much harder — the same settings leave it at just 0.7834 and 0.7713, and pushing sparsity to 0.80 drags it down to 0.7584.
What does an SSIM of 0.76 actually mean? Roughly speaking, the structure and composition are still intact, but texture, detail, and high-frequency information no longer match the original output — visible once you zoom in. Frame-list-to-video drops less because it's often anchored by a first-frame or keyframe constraint; the model is already pinned down by the condition frame, leaving less freedom for sparsification to lose.
Translated into cost, the math gets clearer. On-demand rental pricing for 8x H200 setups runs roughly $2 to $3 per card per hour on the market, putting a full machine at close to $20 an hour (a rough estimate — cloud vendors and contract terms vary considerably). At 1.95x faster, the lossless tier means producing close to double the output in the same hour, cutting per-clip cost to just over half — and that half comes free, without trading away any image quality. The higher tiers save the same money, but they subtract from SSIM to get there, and whether that trade is worth it depends on whether the clips are headed for delivery or for screening.
Laid out this way, practical usage splits into three tiers: use the lossless tier for final delivery, a free 1.95x; use the Cache-DiT conservative tier for balance, 2.65x to 2.99x, with SSIM above 0.90 generally acceptable; use the SubBlock 0.75 tier for sifting through compositions or previews, trading a "roughly legible" image for 6x speed — since anything selected still gets re-rendered anyway.
A few boundaries worth noting
The report draws its own limits. Only three acceleration methods were tested; equally common lossy methods like quantization and progressive resolution weren't included. Timing excludes service startup, warmup, HTTP polling, and file downloads, meaning these multipliers are purely on the inference side — the end-to-end feel will be somewhat less dramatic. SubBlock also comes with hard prerequisites: it only applies to long, non-causal attention, with BF16, a head dimension of 128, and sequence length no shorter than 4096. The test data was actually collected on August 18, nine days before the report was published.
Placed on H3's timeline, the pattern is clear: the model's weights opened in late July, and a month later a third-party inference stack had already pushed the lossless tier to near double speed — an open-weight model's real value is often less about launch-day benchmarks, and more about how many people are willing to write kernels for it once the weights are out.
Sources: LMSYS engineering blog, CocoLoop, SGLang project documentation; speedup multiples and SSIM figures for each tier verified against the same 8x H200, 1344×768, 24 FPS, 50-step denoising configuration.