A new experimental model called 10Eros-Max is pushing MiniMax H3 in an unusual direction. Rather than training a conventional fine-tune from scratch, the project experiments with transferring learned characteristics from other diffusion models directly into selected attention and feed-forward layers of MiniMax H3.
If you’re thinking about purchasing a new GPU, we’d greatly appreciate it if you used our Amazon Associate links. The price you pay will be exactly the same, but Amazon provides us with a small commission for each purchase. It’s a simple way to support our site and helps us keep creating useful content for you. Recommended GPUs: RTX 5090, RTX 5080, and RTX 5070. #ad
The result is an evolving family of MiniMax H3 variants designed to explore how different models can influence H3’s motion, visual style, and generation behavior while retaining much of H3’s original architecture.
The project is available on Hugging Face and currently includes several variants, including hybrid, Turbo, INT8, and FL2VA releases.
What Is 10Eros-Max?
10Eros-Max is an experimental modification of the MiniMax H3 video-generation model.
According to the project’s documentation, the initial approach involved transferring characteristics from Eros 1.5 into MiniMax H3 through cross-attention manipulation. The project later expanded to incorporate patterns from other models, including LTX 2.3, Wan 2.2, and Krea 2.
This makes 10Eros-Max quite different from a traditional LoRA or fine-tuned checkpoint.
Instead of teaching H3 a new concept through a conventional training dataset, the project modifies selected parts of the model’s internal transformer architecture. The goal is to introduce characteristics from other models while leaving important parts of H3 intact.
The author describes the project as an ongoing experiment rather than a finished production model.
How Does the Grafting Work?
The interesting part of 10Eros-Max is its approach to cross-model weight transfer.
The MiniMax H3 architecture contains transformer blocks with attention and feed-forward components. The 10Eros-Max project experiments with modifying selected portions of these components using information from other diffusion models.
In an earlier version, the project mapped attention heads from LTX into corresponding regions of H3. The documentation describes mapping 32 LTX attention heads into a region of H3’s 56 heads and adjusting self-attention behavior.
Later versions expanded the approach considerably.
The project experimented with LTX 2.3 in earlier blocks, Wan 2.2 in middle-to-later blocks, and eventually Krea 2 as an image-model contributor. This is particularly interesting because Krea 2 is an image-generation model rather than a video model.
That means the project isn’t simply trying to make another MiniMax H3 fine-tune. It is effectively exploring whether different models can contribute different “characteristics” to H3 at the architectural level.
Why Is This Interesting?
The idea behind 10Eros-Max is interesting for local AI video generation because it explores an alternative to conventional fine-tuning.
Normally, if you want to change the behavior of a video model, you might train a LoRA or fine-tune the model using a carefully prepared dataset.
10Eros-Max takes a different route.
Instead of training H3 to reproduce a new visual behavior from a dataset, the project attempts to transfer existing learned behavior from other models into specific parts of H3.
This could potentially provide a way to experiment with:
- Motion characteristics from other video models
- Different visual styles
- Changes in how the model responds to prompts
- Different image-to-video behavior
- Alternative generation trajectories
- Hybrid characteristics from multiple models
The project documentation emphasizes that the modifications are targeted rather than replacing the entire H3 architecture.
10Eros-Max Turbo
One of the more interesting recent developments is the 10Eros-Max H3 Turbo hybrid.
The repository currently provides a 10Eros_Max_h3_TURBO-hybrid_beta4.safetensors checkpoint, along with an INT8 version. The full-precision hybrid checkpoint is approximately 40.2 GB, while the INT8 version is approximately 21 GB.
The repository recommends using the Turbo model with a relatively small number of sampling steps. The current documentation recommends six steps with the multires/simple or er_sde/simple scheduler configurations.
For users running MiniMax H3 locally, this is potentially one of the most attractive aspects of the project.
A lower-step Turbo model can significantly reduce generation time compared with a conventional high-step workflow.
Of course, fewer sampling steps do not automatically mean better quality. The actual result depends heavily on the workflow, scheduler, resolution, GPU, and other settings.
INT8 Version
For users with limited VRAM, the INT8 release is particularly interesting.
The current INT8 Turbo hybrid checkpoint is around 21 GB, compared with roughly 40 GB for the standard Turbo hybrid checkpoint.
That doesn’t make the model lightweight, but it makes the checkpoint considerably more practical for high-end consumer GPUs.
For local ComfyUI users, this could be an important consideration when deciding which version to test.
FL2VA and Reference-Based Generation
The repository also includes an 10Eros_Max_h3_fl2va_beta2_pruned.safetensors model.
The checkpoint is approximately 40.2 GB and is designed for the FL2VA branch of the H3 ecosystem.
This is particularly interesting for users experimenting with reference-based video generation, since reference workflows can provide a different way of controlling the subject and visual appearance compared with ordinary text-to-video generation.
However, these releases are still experimental, so users should expect differences in behavior between checkpoints and workflows.
What Can You Expect From 10Eros-Max?
It is important to set expectations correctly.
10Eros-Max is not simply “MiniMax H3 but better.”
It is an experimental model modification, and the results can vary significantly depending on the checkpoint.
The project itself describes the work as evolving and notes that H3 training and the grafting methodology are still being developed.
Some users may immediately notice differences in motion or visual behavior, while others may find the differences relatively subtle depending on their prompts and workflow.
This is also why comparing 10Eros-Max against the original H3 using identical prompts can be more informative than relying on individual showcase images.
Available 10Eros-Max Models
The repository is already becoming a small family of related models rather than a single checkpoint.
Some of the notable releases include:
10Eros_Max_h3_TURBO-hybrid_beta4
The current Turbo hybrid checkpoint, approximately 40.2 GB.
10Eros_Max_h3_TURBO-hybrid_beta4_int8_convrot
An INT8 version of the Turbo hybrid model, approximately 21 GB.
10Eros_Max_h3_fl2va_beta2_pruned
A pruned FL2VA checkpoint, approximately 40.2 GB.
The repository continues to evolve, so users should check the model page for the newest checkpoints and updated workflow recommendations.
Is 10Eros-Max Worth Trying?
If you’re already experimenting with MiniMax H3, ComfyUI, or local AI video generation, 10Eros-Max is definitely worth watching.
The biggest reason isn’t necessarily that it immediately replaces the original H3.
Instead, it demonstrates an interesting research direction: transferring characteristics between different generative models without conventional full-model fine-tuning.
The combination of LTX, Wan, Krea, and MiniMax H3 is particularly unusual. It shows how researchers and independent developers are increasingly experimenting with model architectures themselves rather than treating pretrained checkpoints as fixed products.
For enthusiasts with powerful GPUs, the Turbo and INT8 versions also make the project especially interesting from a practical standpoint.
Final Thoughts
10Eros-Max is still very much an experiment.
Its development shows how rapidly the open-source AI video ecosystem is evolving. Instead of waiting for a new foundation model, developers can now experiment with the internal components of existing models and combine characteristics from completely different architectures.
Whether the approach ultimately produces a consistently superior MiniMax H3 variant remains to be seen.
For now, however, 10Eros-Max is an intriguing project to follow—particularly for anyone interested in MiniMax H3, ComfyUI, model merging, cross-model transfer, and local AI video generation.
The project is available on Hugging Face, where new checkpoints and documentation are being added as development continues.
References
- 10Eros-Max — Hugging Face
- MiniMax H3 — Hugging Face
- LTX 2.3 — Hugging Face
- Krea 2 Raw — Hugging Face
- MiniMax H3 documentation
Leave a Reply