Genmo released Mochi 1 with a stated four-GPU floor

The shot

Genmo's own repository records its first commit, publishing the open weights and code for Mochi 1, on 22 October 2024. Around the same date the company published a blog post, undated on the page itself, titled 'Mochi 1: A new SOTA in open-source video generation models', describing 'a research preview' released 'under the Apache 2.0 license' and stating the model is 'the largest video generative model ever openly released' at 10 billion parameters. The Hugging Face model card, as retrieved on 16 September 2026, confirms the same license and architecture, built on what Genmo calls an 'Asymmetric Diffusion Transformer.'

What the documents show

The blog post states Mochi 1 'dramatically closes the gap between closed and open video generation systems' and reports its own prompt-adherence and motion-quality evaluations, the latter scored with 'Elo scores... computed following the LMSYS Chatbot Arena protocol' against unnamed competitors, a methodology claim rather than a published leaderboard result. The Hugging Face model card corroborates the architecture and license independently of the blog, and adds a detail the blog does not foreground: the model 'requires at least 4 H100 GPUs to run,' a hardware floor that shapes who can operate the checkpoint at all, regardless of its license terms.

The workflow

A team adopting Mochi 1 in this window worked from downloadable weights, a GitHub repository and a free hosted playground at launch, not a production-grade service; the model card's own safety note advises that 'organizations should implement additional safety protocols' before any commercial deployment. The blog states the initial checkpoint generates only 480p video, with a higher-resolution 'Mochi 1 HD' described as a planned, not yet released, follow-up, so a pipeline built around Mochi 1 at this stage needed its own upscaling or finishing step if a higher-resolution master was required.

What the tool does not change

The model card lists its own limitations candidly: 'in some edge cases with extreme motion, minor warping and distortions can also occur,' and the model 'does not perform well with animated content' because it was optimized for photorealistic styles. Those are the developer's own stated boundaries, not an outside critique, and they mean a production still needs a human review pass to catch warping or unwanted stylistic drift before a generated clip is cut into a sequence.

• Does a stated Elo or preference score come from a published leaderboard, or from the developer's own internal comparison.
• What hardware floor does the model card state, and does a production's infrastructure actually meet it.
• Is the checkpoint under evaluation the base research preview, or a later higher-resolution variant with different stated limits.

Genmo's own repository, blog post and model card together date and describe a specific research-preview checkpoint, openly licensed but bounded by a stated hardware requirement, a 480p ceiling and named failure modes the company disclosed itself.