Mochi 1

Mochi 1 is Genmo's open-weight text-to-video model, released as a research preview on October 22, 2024 under the Apache 2.0 license. At 10 billion parameters, Genmo billed it at launch as the largest video generative model ever released openly, and it remains the reference point for what a fully open, commercially usable video diffusion model looks like.

Overview

Genmo is a San Francisco research lab founded in 2022 by Paras Jain and Ajay Jain, both UC Berkeley PhDs - Jain co-authored some of the foundational diffusion model papers (DDPM-adjacent work, Dream Fields, DreamFusion) before starting the company. Mochi 1 was Genmo's first public open-source release and landed the same week as its Series A funding round, positioning the model as a credible open alternative to closed video generators like Runway, Kling, and Google's Veo line.

The model ships as a research preview, not a finished consumer product: 480p resolution, 5.4-second clips, and a hardware floor that starts at a single H100. Genmo has never operated a first-party paid API for Mochi 1 - anyone who wants to run it goes through self-hosted infrastructure, a community ComfyUI workflow, or third-party inference hosts like Replicate and fal.ai that wrap the open weights. Our video generation pricing comparison tracks Mochi 1 in that category: no managed rate card, just the cost of the GPU-hour it runs on.

Key Specifications

Benchmark Performance

Video generation lacks a single agreed-upon benchmark the way LLMs have MMLU or GPQA. The closest thing is VBench, which our own video generation benchmarks leaderboard tracks across open-source releases - but Mochi 1 has never had a full VBench aggregate score published, so the honest comparison is a mix of what is measured (resolution, license, hardware) and what Genmo itself has claimed.

Genmo's launch claims should be read as self-reported: the company assessed prompt adherence with an automated Gemini judge and motion quality with human Elo voting modeled on LMSYS Chatbot Arena, then concluded Mochi 1 performs "very competitively with the leading closed models." Independent third-party writeups have echoed the motion-quality claim - Mochi's 30 FPS output and strong physical coherence are consistently the model's most-cited strength - but nobody has published a controlled Mochi-vs-Wan-vs-HunyuanVideo VBench run using the same protocol. Treat the comparison as directional, not a settled scoreboard. Where Mochi 1 unambiguously loses ground is resolution: 480p in July 2026 is well behind Wan 2.2's 720p+ and HunyuanVideo 1.5's native 720p, let alone Wan 2.7 or LTX-2.3's 4K ceiling.

Key Capabilities

Motion physics over polish

Mochi 1's most consistent reputation, across Genmo's own benchmarking and independent reviews, is realistic motion - fluid physics, natural object interaction, and fewer of the warping artifacts that plague competing open models at fast camera movement. The AsymmDiT architecture allocates roughly four times more capacity to the visual stream than the text stream (3072 vs. 1536 hidden dimensions), a design choice Genmo made specifically because video quality, not text comprehension, was the harder problem to solve.

Fully open weights, no vendor lock-in

Apache 2.0 covers the full stack - inference code, model weights, and the training recipe outline - for personal and commercial use with no attribution requirement. That's a meaningfully more permissive license than HunyuanVideo 1.5's custom terms, and it's why Mochi 1 became a base model for LoRA fine-tuning projects within weeks of release: Genmo shipped official LoRA support in November 2024, letting teams adapt the model to specific visual styles or datasets without touching the base weights.

Consumer-GPU accessibility via ComfyUI

The 60GB VRAM requirement for the reference bf16 implementation puts Mochi 1 out of reach for consumer hardware out of the box. ComfyUI's community integration closed that gap with fp8-scaled text encoder and diffusion model variants that fit under 24GB VRAM - enough to run on a single RTX 4090. That's the deployment path most individual creators actually use; the H100-class self-hosting numbers below apply to teams running it at scale or through serverless GPU platforms.

Pricing and Availability

Genmo doesn't sell a metered API for Mochi 1. There's no per-second or per-token rate card because there's no first-party commercial endpoint - Mochi 1 is a self-hosted model by design. That puts it in a different cost bucket than the API-priced competitors on our AI video generation pricing page, where every other entry has a published per-second rate.

Three practical access paths exist:

1. Self-host on rented GPU infrastructure. Modal's published example puts a single generation at roughly $0.33 on an H100 at about $5/hour, though the docs note a clip can take several minutes of compute. For teams already comparing GPU rental math, our open-source hosting costs page has current H100 rates across RunPod, Lambda Labs, and other providers - Mochi 1's per-clip math follows the same hourly-rate-to-output-time formula used there for LLM tokens.
2. Third-party inference hosts. Replicate lists Mochi 1 at roughly $0.42 per run on H100 hardware; fal.ai prices it around $0.40 per video. Neither is Genmo's own infrastructure - these are community/partner wrappers around the open weights, priced independently.
3. Local or on-prem GPU. With ComfyUI's quantized build, a single RTX 4090 (under 24GB VRAM) can run Mochi 1 with no per-generation cost beyond electricity and hardware ownership - the appeal for hobbyists and small studios that already own the hardware.

There's no free tier or trial credit system to speak of beyond whatever a third-party host offers, since Genmo isn't running a consumer product business around Mochi 1 the way ByteDance runs Dreamina or Google runs Veo through Google Vids.

Strengths and Weaknesses

Strengths

• Apache 2.0 license covering weights, code, and commercial use with no attribution requirement
• Strongest self-reported motion-physics quality among open video models at launch, and still frequently cited for fluid motion
• Official LoRA fine-tuning support since November 2024 for style- or dataset-specific adaptation
• Runs on a single H100, or under 24GB VRAM with ComfyUI's fp8-quantized build - no data-center minimum
• Full 3D attention over 44,520 visual tokens gives coherent long-range motion within its clip length

Weaknesses

• Capped at 480p - the promised 720p "Mochi 1 HD" upgrade was never shipped, and GitHub issue #132 tracking it has no release date as of July 2026
• 5.4-second maximum clip length is short next to Kling 3.0's 4K/60fps outputs or Seedance 2.5's 30-second clips
• No native audio generation, unlike Veo 3.1 or Seedance 2.0
• No managed API from Genmo - every access path routes through self-hosting or a third party, which adds setup friction most closed-model users don't face
• No published VBench aggregate score, so head-to-head quality claims rest on Genmo's own launch-day methodology rather than an independent, repeatable benchmark

Related Coverage

• AI Video Generation Pricing Comparison - where Mochi 1's lack of a managed API sits against Veo, Kling, Runway, and Seedance per-second rates
• Open Source LLM Hosting Costs - the GPU-hour cost formula behind Mochi 1's self-hosting math
• Video Generation Benchmarks Leaderboard 2026 - full VBench and Elo standings, including Mochi 1's entry in the open-source rankings
• Best AI Models for Video Generation - where Mochi 1 fits against the current Elo leaders
• LTX-2.3 Model Card and Wan 2.7 Model Card - the open-source models that have since overtaken Mochi 1 on resolution and speed

FAQ

Is Mochi 1 free to use?

The weights are free under Apache 2.0 for personal and commercial use. Running it isn't free unless you already own a capable GPU - expect to pay for GPU rental (self-host) or a per-generation fee at a third-party host like Replicate or fal.ai.

Does Genmo offer a Mochi 1 API?

No managed API from Genmo. Third parties like Replicate ($0.42/run) and fal.ai ($0.40/video) host the open weights independently, and self-hosting on Modal runs about $0.33/generation on an H100.

What resolution does Mochi 1 support?

480p only. Genmo announced a 720p "Mochi 1 HD" version for later 2024, but it has never shipped as of July 2026.

Can Mochi 1 run on a consumer GPU?

Yes, with caveats. The reference implementation needs about 60GB VRAM on an H100-class GPU, but ComfyUI's fp8-quantized build brings that under 24GB, enough for a single RTX 4090.

How does Mochi 1 compare to Wan 2.2 or HunyuanVideo 1.5?

Mochi 1 has no published VBench aggregate score, while Wan 2.2 scores roughly 84.7% and HunyuanVideo 1.5 posts 96.4% visual quality / 68.5% text alignment on VBench sub-metrics. Mochi 1's cited strength is motion physics rather than resolution or a comparable benchmark number.

Sources

• Mochi 1: A new SOTA in open text-to-video - Genmo Blog
• genmo/mochi-1-preview - Hugging Face
• genmoai/mochi - GitHub
• GitHub Issue #132: HD 720p Release timeline
• Genmo introduces Mochi 1, an open-source text-to-video generation model - SiliconANGLE
• AI video startup Genmo launches Mochi 1 - VentureBeat
• Genmo Mochi 1: A new benchmark for open AI video models - The Decoder
• Run Mochi in ComfyUI with consumer GPU - ComfyUI Blog
• Text-to-video generation with Mochi - Modal Docs
• Mochi 1 by Genmo - Replicate
• Mochi 1 - fal.ai
• Genmo - About