--- title: SOL-H3 Niu Lai Promo emoji: 🐂 colorFrom: yellow colorTo: yellow sdk: static pinned: false --- SOL-H3 promotional video gallery from the NVIDIA Sana team. The featured native 16:9 high-bitrate edition is a freshly regenerated 33.875-second SOL-H3 performance story built around the audited 8× NVIDIA B300 results published on the official NVIDIA Labs page. At 1344 x 768 and 24 FPS with four DiT forwards, the median request latencies shown in the video are 1.669 seconds for 5 seconds / 124 frames, 3.738 seconds for 10 seconds / 243 frames, and 6.619 seconds for 15 seconds / 362 frames. The release uses a clean opening-audio prompt contract so the first humanlike sound is the complete line “Video generation is slow?”. The on-screen label remains exactly SOL-ATTN, while its spoken form is explicitly “Sol Attention”. The story generates no acknowledgement scene and moves directly from the benchmark to the NVIDIA and SANA / SOL-ENGINE finale. The H.264 master uses CRF 10 and raises the continuous electronic background bed by 2 dB versus the previous master. A new experimental 33.875-second LightX2V Ref2VA Turbo 8-step v1.0 768p edition uses the same prompts, four references, seeds, frame counts, and mastering as the 4-step edition. It runs eight effective DiT evaluations from nine scheduler grid points at the checkpoint-native video/audio shifts 6/3. The official LoRA file is minimax_h3_ref2v_turbo_8step_v1.0_768p_bf16.safetensors (SHA-256 9bac880b1a5d7ac052171cf6cce769f0cceaaa42ffa51de4b8e41143a2bdd2d2). Visual QA retained the raw model output for honest comparison: the FastH3 subtitle omits the leading 4-, while SOL-ATTN, the audited benchmark board, AND MORE, and the NVIDIA plus SANA / SOL-ENGINE finale remain clear. A separate freshly regenerated 33.875-second quality-comparison edition runs the same SOL-H3 story and four references with the official MiniMax-H3 Ref2VA base model and no LoRA. Its official num_inference_steps=50 scheduler contains 50 grid points and performs 49 DiT evaluations because the terminal zero is not evaluated. It applies the same clean opening-audio and “Sol Attention” pronunciation contract, direct no-acknowledgement story structure, continuous +2 dB background bed, exact 813-frame duration, and CRF 10 high-quality mastering. The hero is a complete 30-second Niu Lai advertisement. It retains the established ten-second introduction and adds two newly generated LightX2V Ref2AV Turbo 4-step segments. The capability sequence uses three clean pages with at most two oversized cards on screen at once, followed by a single `AND MORE...` card. Its narration is shorter and more relaxed, while the exact performance board and the NVIDIA plus two-level SANA / SOL-ENGINE lockups remain inside the video itself. A separate 34.625-second native 1344 x 768 high-bitrate landscape edition adds the four-step FastH3 source model from Hao AI Lab at UCSD to the AI-native SOL-Attn stack reveal. Its H.264 video stream measures about 10.25 Mbps, while the updated B200 board shows 5-second video in 2.625 seconds, 10-second video in 6.839 seconds, and 15-second video in 12.884 seconds. It is displayed independently; the portrait version and original gallery videos remain available. The gallery additionally retains the existing 20-second LightX2V Ref2VA and dense Fast H3 I2V 4-step comparisons, plus the 50-step Ma Lai and Kapi Lai animal variants.