Hey Herd,

It's been awhile since we've done a proper feature update, so wanted to come up for air to show you the latest and greatest. The Oxen team has been heads down plowing the fields, building features you've asked for. It's now time to harvest the crops.

The highlights:

• A unified workbench for all your models
• Brainstorm with LLMs in the workbench
• Prompting @image1 autocompletion
• Uncensored access to Seedance 2.0 4k + many new models

As always, feel free to let us know if you have any feedback!

A great way to get in contact is our Discord or by emailing hello@oxen.ai directly. We're always looking to improve the experience.

AI Workbench

All your models, in one place. If you haven't tried our model workbench lately, I am sorry to inform you, you are behind. Just kidding! Nothing like a light dose of AI anxiety to start an email. You're fine. Feel free to take it slow, at your own pace, we will all learn the right tools when we need them. That being said, Oxen.ai is slowly becoming many artists and engineers favorite place to build and create.

Did we set out to build a Higgsfield competitor? Not really.

And I wouldn't say that's what this workbench is. Yes it is a model aggregator, but it is also the data storage layer, the provenance layer, the versioning layer, all built into one. It's meant for big teams and professional users. It's as if your favorite model aggregator grew up and had a baby with GitHub, Airtable, and Dropbox. Generate your assets, track and version them, organize them into beautiful databases, and scale to terabytes. More on that later in "The Core".

If you haven't seen what the latest interface looks like, here's a quick 1 minute video demo:

We also have very transparent pricing and don't play games to try to get you to subscribe. Oxen are trustworthy, straight forward, simple creatures. Here to do work, and not play games.

Brainstorm with LLMs

One feature of the model workbench that we added recently was the ability to brainstorm with your favorite LLMs right next to all your creations. Quickly switch to your favorite Claude or GPT, drag and drop images into it, and start brainstorming.

@image1 Autocomplete for Prompting

Many of you gave us feedback that it was a pain to connect your reference images to your prompts in the workbench. So we fixed it. Now you can type in @image1 or @image2 and we will magically translate this to the correct syntax for the model on the backend.

The Core

The core open source version control infrastructure is getting stronger by the day. So strong that we are getting adopted by larger organizations like the ones in the Motion Picture Laboratories, Inc. (MovieLabs) initiative to help teams collaborate more efficiently from production through post and VFX. MovieLabs is a collaborative technology roadmap developed by major studios (Disney, Paramount, Sony, Universal, and Warner Bros.) that has a 2030 vision that we are helping bring to live with our version control system. Provenance, versioning, software-defined workflows, security, and end-to-end tracking of assets are all key pillars of the vision.

On our roadmap for the core VCS layer is supporting different filesystem backends like S3, better compression and deduplication of data, and continuing to make our core merkle tree faster. If you enjoy this type of low level optimization work, we are hiring! Send us your resume via email to hello@oxen.ai. Tell us about the type of projects you like to work on, and we will get back to you ASAP.

The Models

The best place to start is poking around the latest models we've added. Some of our favorite from the past month are Gemini Omni Flash, Seed Audio, GLM 5.2, Krea 2, and of course Seedance 2.0 4K (uncensored). You can wire any of these models up to your app via API, or just use them directly in our workbench.

Explore the full model list here. I'll do a detailed breakdown below.

Seedance 2.0 4K: Uncensored & Unfiltered

Seedance 2.0 is by no means a new model, but it is still by far the best video generation model that is available right now. What they did add recently is a 4K option for generation. This means you do not need to use another model for upscaling after the generation is done, the model does it natively.

At Oxen.ai, we also have special access to the model where we turned off all the filters for you. No more "this content is not allowed" or rejecting of human faces. I had a blast playing around with it to see how close it could get to "Kill Bill" style cinematic blood and violence in this unhinged YouTube video below. Watch at your own risk (or enjoyment 😄).

Seed Audio

ByteDance continues to impress with their models. If you haven't played around with audio generation yet, Seed Audio 1.0 is a great place to start. It is a foundation model that can go from any text prompt or audio reference, and generate new audio. This could be people's voices, the sound of rain, or ox hoofs on a dirty path.

It can even do short music clips or jingles for your company, are they any good? This one was not. But a let your prompt engineering flag fly. All it takes is $0.02 and 17 seconds of your time to find out for yourself 😏.

Gemini Omni Flash

This latest video generation model from Google still feels like it lags behind Seedance 2.0 on cinematic quality - but where it does shine is in video-to-video edits and specific use cases like typography and effects that used to require a full post-production workflow.

Here are a few example videos from the model, starting with some kinematic typography:

Then a product video of our new flagship OXEN.AI rain jacket. The team is unclear when manufacturing starts, just venmo me if you want one.

Krea 2

One of the more exciting open source releases in the past few months. Krea pre-trained their own model from scratch, built around taste, style control, and art direction. You can guide this model through aesthetics, style transfer, mood boards, and creative control rather than only literal prompt following. The model includes variants like Krea 2 Raw and Krea 2 Turbo, with the Raw model positioned for fine-tuning and LoRA training and Turbo for fast inference.

Say you had a very specific artistic style of red marker and pen drawings for your brand. You can use Krea 2 Large in our workbench to guide the model with this style reference and generate new images with very simple prompts. In this case our prompt is just "a stadium of fans cheering", and the art direction is done by the reference image.

This model feels much more like a Midjourney --sref than Nano Banana 2 or GPT Image 2.

GLM 5.2

People are loving GLM-5.2 because it hits the sweet spot developers actually care about: near-frontier coding and agent performance, a usable 1M-token context window, it is open-weights and fine-tunable, and much lower cost than closed U.S. frontier models (looking at you Anthropic and OpenAI). It is a beast of a Mixture of Experts model coming in at 753B total parameters, and with roughly 40B active parameters per token. This means about 1.51 TB for the full official weights, not for the faint of heart. You can try GLM 5.2 through our UI or API and wire it up to Aider or Cline.

We also unofficially have released and open sourced our own Oxen Harness. The Oxen Harness comes with both a user interface and a terminal interface. It's also hackable and extensible. All you have to do is wire up your Oxen API key. It also works with local models if your laptop is souped-up enough.

It's also Oregon Trail themed. Obviously. Feel free to check it out here, it's raw, but fun to kick around and hack on yourself. https://github.com/gschoeni/Oxen-Harness

👋

That's it for today! If you ever need a new model added to the workbench, or have ideas for how the product could be better, we are all ears. If you made it this far, you deserve a little prize. Email us at hello@oxen.ai your use case, reference this post, and we'll give you $100 of free credits to try out the latest and greatest.

Happy creating 🐂

~ Greg & The Oxen.ai Herd