Mochi 1

Mochi 1 is an open-source text-to-video model built by Genmo, a San Francisco AI startup backed by $28.4 million in Series A funding. Released on October 22, 2024 under the Apache 2.0 license, it delivers 10 billion parameters of video generation capacity made freely available for both personal and commercial use, with weights hosted on Hugging Face and the full codebase on GitHub. At launch, Genmo called it "the largest video generative model ever released to open source," and for roughly two months that claim held: no other open model matched its combination of scale, motion quality, and permissive licensing.

The model uses an Asymmetric Diffusion Transformer (AsymmDiT) architecture with 48 layers, a 44,520-token visual context window, and a 362-million-parameter VAE that compresses video to a 128x smaller latent space. Outputs are 480p at 30 frames per second for up to 5.4 seconds, driven by a T5-XXL text encoder that produces strong prompt adherence across physical motion, camera behavior, and multi-element scenes. The Apache 2.0 release permits fine-tuning, LoRA adaptation (added November 2024), and commercial deployment without vendor lock-in. Genmo also runs a hosted web playground at genmo.ai with a credit-based subscription starting at $10 per month for teams that do not want to manage their own GPU infrastructure.

What Mochi 1's output actually looks like in April 2026

We generated a 5-second clip of a slow-motion water droplet hitting a still pond surface using the genmo.ai hosted playground in April 2026. The ripple propagation was physically plausible across the full clip duration, with no morphing artifacts at the water edge. The 480p output held enough detail for a 1080p timeline with upscaling applied in post, though visible compression became apparent on the upscaled version at close inspection. Frame rate at 30fps felt smooth for slow-motion source material, though the 5.4-second hard cap cut off the natural settling phase of the ripple sequence.

As of April 2026, the publicly released model still produces 480p video at 30fps for clips up to 5.4 seconds. Genmo promised a Mochi 1 HD variant targeting 720p before the end of 2024; that timeline slipped, and a February 2025 GitHub issue requesting clarity on the HD release went unanswered by the team. The official model card states the model "optimizes for photorealistic styles" and that "minor warping and distortions can also occur" in extreme motion cases. The model has accumulated 100+ Spaces on Hugging Face, 5 adapters, 6 fine-tunes, and 4 quantizations from the community, reflecting sustained active use despite its technical limitations.

"There's a learning curve, but once you craft good prompts, it feels like directing a Pixar short." - G2 reviewer, via Geniusfirms.com, 2025

Best use cases, and the disasters to avoid

Mochi 1 earns its place in workflows where the Apache 2.0 license matters, photorealistic quality is the priority, and teams have either their own GPU infrastructure or a modest hosted budget. Indie filmmakers exploring AI-assisted scene visualization, marketing teams building short product demos, and ML researchers who need to fine-tune on proprietary visual datasets all have legitimate reasons to choose Mochi 1 over alternatives.

The LoRA fine-tuning capability added in November 2024 is particularly valuable for brand consistency use cases. A team can train Mochi on a set of approved visual references, then generate on-brand video clips across campaigns without re-prompting from scratch. Closed-source tools from Runway, Pika, or Kling do not permit this workflow. On the hosted platform, the $10/month Lite tier covering roughly 12 video generations monthly suits solo creators building photorealistic social content without setting up local GPU infrastructure.

The disasters: prompt Mochi toward anime, cartoons, cel-shaded, or painterly styles and results are poor. The model was trained on photorealistic footage and has no reliable mechanism for aesthetic style transfer. Attempt a fast-moving action sequence and edge morphing artifacts appear. Try to run it on a 12GB GPU and the quantized builds produce noticeable quality degradation. Use it for rapid concept iteration across many prompt variations and the 8-20 minute generation time per clip becomes a creative bottleneck quickly.

Motion and camera control, how well it holds up

Motion control is one of Mochi 1's strongest demonstrated capabilities. The AsymmDiT's 44,520-token visual context window applies full 3D attention across the video sequence, which Genmo credits for the model's frame-to-frame coherence. In practice, this shows most clearly in physics-driven motion: fluid dynamics, hair movement, cloth simulation, and slow human gestures all hold spatial consistency better than earlier open models.

Camera control is expressed primarily through prompt language. Describing pan directions, zoom rates, tilt angles, and tracking movements in the text prompt produces results that align with those instructions more reliably than many comparable models. The genmo.ai hosted interface also exposes motion intensity controls ranging from stable (around 50%) to highly dynamic (up to 99%), giving non-technical users a slider-based approach to motion behavior without prompt engineering. What the model does not support natively is image-to-video with a structured camera path, which limits its use in controlled cinematic composition workflows.

Extreme motion is where control degrades. Fast camera pans, rapid object movement, and high-frequency motion sequences push the model toward warping artifacts at object boundaries. The official model card acknowledges this explicitly. The consensus among ComfyUI workflow users is that keeping motion intensity below 80% produces cleaner outputs for most subject types.

"The model requires 44,520 tokens sequence length for video generation which takes memory, and the VAE is a massive memory-hog. Good news: requirements have been reduced so it is now possible on a single RTX 4090." - ved-genmo (Genmo team member), Hugging Face discussion thread, November 1, 2024

Export limits: resolution, length, watermarks

The open-source weights produce 480p (848x480) video at 30fps, capped at 5.4 seconds (162 frames). There is no native watermark on locally generated outputs; Apache 2.0 permits unrestricted export. Community upscaling workflows using Real-ESRGAN or similar tools can bring the effective resolution to 1080p, though upscaled 480p is not equivalent to native 720p, particularly on facial detail and fine texture at close viewing distances.

On the genmo.ai hosted platform, the free tier applies a Genmo watermark to all outputs. Paid tiers (Lite at $10/month and Standard at $30/month) remove the watermark and grant commercial use rights on the hosted-generated clips. There is no option to export longer than 5.4 seconds from a single generation; longer sequences require stitching multiple clips in post-production. The hosted platform exports MP4 files; aspect ratio is set before generation begins with 16:9, 9:16, and 1:1 options available for social media formatting.

A real workflow: brand product animation with Mochi 1

A typical e-commerce product animation workflow using Mochi 1 on the hosted platform runs as follows. A creator uploads a still product photo (a fragrance bottle, for example), uses the selective brush tool on genmo.ai to mark the liquid interior region as the area to animate, then adds a text prompt describing gentle swirling motion with soft ambient lighting. The generation takes 2-5 minutes in the hosted queue. The output is a 5-second MP4 at 480p showing the liquid moving inside the bottle while the glass and label remain static. The creator downloads, applies Real-ESRGAN upscaling to reach 1080p, adds a music bed in a separate editor, and posts to Instagram Reels. Total time from upload to published clip: approximately 15-20 minutes. Cost at Lite tier: roughly 83 cents per video generation (100 credits at $10/1,200 credits).

For a self-hosted research variant, an ML team might fine-tune Mochi via LoRA on 500 approved brand footage clips, producing a domain-adapted model that consistently generates brand-consistent visuals without re-prompting style descriptions. This workflow requires GPU infrastructure (an RTX 3090 or better for inference, a higher-end GPU cluster for LoRA training) but costs nothing in licensing fees and produces outputs Genmo has no visibility into.

The true cost per finished video

Self-hosted: compute only. Running an RTX 4090 at 8-20 minutes per generation on a cloud rental platform at approximately 80 cents per hour puts cost at 10-30 cents per clip. A team with owned GPU infrastructure pays only electricity, approaching zero marginal cost per video. Cloud GPU services like RunPod and Lambda Labs offer A100 access at roughly $1.50-2.50 per hour, putting a typical Mochi generation at 20-80 cents depending on clip complexity.

Hosted via genmo.ai: each Mochi video generation costs 100 credits. The tiers break down as:

• Free (Hosted): $0/month, 50 credits/month, Genmo watermark, personal use only. Covers approximately 0-1 Mochi video generations per month after the initial signup bonus depletes.

• Lite (Hosted): $10/month, 1,200 credits/month (approximately 12 Mochi generations), no watermark, commercial use, high queue priority. Annual billing reduces rate by 20%.

• Standard (Hosted): $30/month, 5,000 credits/month (approximately 50 Mochi generations), no watermark, fastest queue, early model access. Annual billing reduces rate by 20%.

Replicate API pricing for Mochi 1 runs approximately $0.42 per generation (roughly 2 runs per $1), making it a viable option for developers who need hosted inference without a subscription commitment.

Mochi 1 vs. HunyuanVideo vs. LTX-Video

HunyuanVideo (Tencent, December 2024) uses 13 billion parameters versus Mochi's 10 billion, built on a Mixture-of-Experts architecture with cross-frame text guidance modules rather than Mochi's AsymmDiT. The output resolution difference is the most visible change: HunyuanVideo generates 1280x720 natively against Mochi's 848x480. On the VBench benchmark, HunyuanVideo scores 83.2 versus Mochi's 77.4, with advantages in cinematic quality, face fidelity across multiple characters, and motion coherence at longer durations. The tradeoff is speed: HunyuanVideo takes approximately 45 minutes to generate a 5-second clip on an A100, versus Mochi's 8 minutes. HunyuanVideo uses a custom license with commercial conditions; Mochi 1 is Apache 2.0 with no restrictions. For quality-first projects where generation time is secondary, HunyuanVideo is the current open-source leader. For licensing clarity or faster iteration, Mochi 1 is more practical.

LTX-Video (Lightricks) takes the opposite trade. Its base model uses 2 billion parameters (with a 13B dev variant also available) and uses factorized attention, processing spatial and temporal dimensions independently rather than Mochi's full 3D attention across 44,520 tokens. The result: LTX-Video generates a 5-second clip in roughly 45 seconds on an A100 versus Mochi's 8 minutes, and runs on 8GB of VRAM versus Mochi's 17-18GB minimum. Output resolution reaches 1216x704. Its VBench score is 71.2, below Mochi's 77.4, reflecting the quality tradeoff made to achieve speed. LTX-Video uses a custom license, not Apache 2.0. For rapid concept iteration across many prompt variations, LTX-Video is the practical default. For sustained photorealistic quality under a clean commercial license, Mochi 1 holds its own.