01
Accelerate
State-of-the-art inference optimization, tuned to your hardware and workload, so every video costs less and arrives sooner.
Accelerated video models.
Your data. In production.
Self-improving.
Full-stack open-weight video models, from data to production. Paired with agents for creative workflows and data-driven self-improvement.
Want to have your own real “unlimited” video generation on your data?
FastH3 on your dedicated host. Your data, dedicated capacity, one flat hosting cost.
≈$0.014per second of 768p video
or≈4,000768p videos per day
- 82% cheaper than MiniMax H3 API
- 90% cheaper than Seedance 2.0 API
Or feeling ambitious for real-time generation?
Illustrative GPU cost at full utilization: 8× RTX PRO 6000 × $3/hour/GPU; 10-second 768p videos with audio. Excludes service fees. API comparison: H3 768P $0.08/s; Seedance 2.0 720p $0.15/s (Sep 30, 2026).
What We Do
From open weights to production. Better with every iteration.
02
Customize
Adapt open models to your visual language, formats, and workflows while preserving quality and creative control.
03
Deploy
Move from weights to a dependable production service with scalable infrastructure, observability, and hands-on support.
04
Integrate
Connect video generation to your business systems through agents that generate assets, organize your library, and execute your production workflows.
05
Self-improve
We keep your deployment up to date with model and inference upgrades. Connected to your business systems, we use production feedback to refine agent strategies and improve end-to-end quality, speed, and cost.
Open Weights
We believe open video models will remain a critical part of the production stack. The Nuva Lab team has contributed to open-source AI for years. We optimize open-weight models for production today and will continue supporting and contributing to the models that come next.
FastH3
Current Work
MiniMax H3, post-trained and optimized by FastVideo.
Sparsity Kernel
T2V Validated
Omni Ref Pending
90% Sparse≈10× fewer target-video QK/PV block pairs
Few-Step Optimization
In Production6.25× Fewer Steps50 → 8 sampling steps
Optimized Serving Runtime
In ProductionLatency and ThroughputFull-host optimization
NVFP4 Native
OptionalBlackwell Native PrecisionBuilt for NVIDIA Blackwell’s native FP4 acceleration
Releases & News
Open model work, production integrations, and ecosystem releases from Nuva Lab and our collaborators.
Optimized FastH3 Serving with Uniserve
LatestAn optimized serving runtime for FastH3 on NVIDIA Blackwell. A 10-second clip with audio renders in 6.3 s on eight GB200 GPUs (5.4 s with optional NVFP4), with 1.2–1.7× lower median latency and 15–45% higher throughput than other open-source serving stacks. Already serving production traffic at Nuva Lab.
FastH3 8-Step V2
The eight-step FastH3 checkpoint that production serving now runs on, trained with data-free DMD2 distillation and 80% sparse video attention.
FastH3 Goes Local
The FastVideo ecosystem brings FastH3 to NVIDIA DGX Spark and Apple Silicon for local video generation.
MiniMax H3 on vLLM-Omni
vLLM-Omni details the complete MiniMax H3 serving stack and its integration with FastVideo's four-step FastH3.
FastH3 Preview v1
Open-weight, four-step sparse-distilled MiniMax H3, developed with FastVideo and NVIDIA collaborators and grounded in production workloads with Nuva Lab.
Partner
Collaborators
Want to have your own real “unlimited” video generation on your data?
Flat cost. Dedicated to you. FastH3 on your dedicated host.