MiniMax H3 native 1440p on a single RTX 4090 via auto-scheduled sparse attention, open-sourced
This_Temporary_8537 · reddit · 2026-08-30
An indie developer open-sourced X-MinimaxH3, adding an automatically scheduled sparse-attention system for MiniMax H3 that achieves native 720p→1440p second sampling on a single RTX 4090: 112s/223s/334s for 5/10/15-second videos (exploratory runs, not tuned limits).
The core is the scheduler: instead of one fixed sparsity ratio across all denoising steps and layers, it assigns per-location attention budgets along the trajectory, staying conservative around regions critical for motion, consistency and fine detail. Users get a single 0–100 acceleration dial, with 0 as the full Dense reference. The 1440p stage is native H3 latent-space second sampling reusing video/audio latent state — not frame-by-frame upscaling.
The project also ships FL2VA, multi-reference Ref2VA, Base/Turbo LoRA switching, a Web UI, REST API and four ComfyUI workflows; the author seeks feedback from consumer-GPU users (3090/5060/4060).
More from Infra
- MiniCPM-V 4.5 MoE reportedly hits 13+ tokens/s on iPhone with just 3GB memory — heyshrutimishra · 2026-10-16
- Solo-built Vulkan inference engine LaForge hits ComfyUI-level SD1.5 speed: 512px in 0.8s, 402MB container — Ichidan256 · 2026-10-16
- Trading 2 RTX 5090s for 4 R9700s: does doubled VRAM pay off? — poofph · 2026-10-16
- Running PolyStrata + Cline coding agent on a $0-fancy laptop at ~90 tok/s — rpwoerk · 2026-10-16
- TensorFold 1.0.12 runs 120B vision model on 1-2 DGX Sparks, adds Mac image input — HankYeomans · 2026-10-16
- vLLM brings frontier model local deployment to NVIDIA RTX Spark PCs — vllm_project · 2026-10-16