There was a talk about Venice AI in a couple of comments, exactly five months ago. For a reason I can’t even explain to myself, I revisited this useless AI offering.

Who are they? And where could you meet them?

According to their ToS page, they are

Venice.ai, a Wyoming company with its registered office at: 1309 Coffeen Avenue Suite 14343, Sheridan, Wyoming 82801, United States.

Except that there isn’t any such thing there. Here’s what can be found at the address: a parcel dropout. No offices. For the record, here’s a view of Suite 1200.

You need to understand that, on paper, Suite 1200 does exist there, while Suite 14343 doesn’t. But that’s OK, because there isn’t any real company there, despite more than 100,000 companies listing 1309 Coffeen Avenue, Sheridan as their address.

Wyoming is a tax shelter state, and Sheridan is the bogus LLC capital of the world. There are two addresses in that town that, together, host hundreds of thousands of companies:

  • 30 N. Gould St., Sheridan, WY 82801
  • 1309 Coffeen Avenue, Suite 1200, Sheridan, WY 82801

It’s not always Suite 1200, but the mail gets delivered anyway. HERE WeGo lists there Embles LLC, but at Suite 5343. According to Google Maps, the most popular places at this address are:

Also at Suite 1200:

But the most relevant is the title of this page: Is your business one of the 100K+ businesses registered at 1309 Coffeen Ave, STE 1200, Sheridan, WY 82801?

Let’s list with some press articles:

Registered Agents, Inc., located at 30 N. Gould St. in downtown Sheridan, represents an estimated 53,267 businesses, according to the Secretary of State Office.

Registered Agents, Inc., though, is just one commercial registered agent listed at that address. According to a roster on the Secretary of State website, a total of 21 commercial registered agents have offices at 30 N. Gould St.

For example, Northwest Registered Agent Service, Inc. represents approximately 5,872 businesses, Registered Agent Service LLC represents 194 businesses and Wyoming Registered Office LLC represents 99.

And nobody answers the phone.

The small building [on North Gould Street in Sheridan] is home to the Sheridan office of Registered Agents Inc., a national corporate services firm; more than 266,000 companies incorporated using the address of this modest office between 2019 and 2024, according to an International Consortium of Investigative Journalists analysis of information provided by the data firm OpenCorporates.

Wyoming’s trust laws give business entities some of the nation’s strongest privacy protections, allowing complete anonymity to their ownership. That’s why one address [30 N. Gould St.] in Sheridan is listed as the place of business for about 120,000 Wyoming LLCs, and why one county assessor says LLCs can ignore her requests for basic business information.

Of course they are.

➡️ So Venice.AI is a company registered at one of the two addresses in Sheridan, Wyoming, that are the front of hundreds of thousands of shell companies that have no real address and no assets! If in Gould St. there are 21 agents that helped create such companies, I don’t know how many there are at 1309 Coffeen Avenue, because in addition to Suite 1200, the fictitious Suites 5343 and 14343 are also declared—but maybe there are more of them!

Would you trust your money with such a company that doesn’t really need to exist as long as the authorities can’t legally know anything about it?

OK, I’ll suppose you’re interested in the free tier, then.

The free tier

The pricing page lists the free offer with a small error, but they have a history of contradicting themselves:

  • Base AI models — I’ll list them later.
  • 25 text prompts per day — True.
  • 15 image prompts per day — Real value: 16 images.

The limit counter resets daily at 12:01am UTC.

This is not particularly enticing, but I suppose one could use such a free service. So I’m going to ignore the API calls, the VVV tokens, and all that shit that was supposed to make Venice.AI attractive. Not with a shell company!

The mobile app isn’t a true app but a PWA (Progressive Web App).

The only real thing that recommends Venice.AI is that you can change the Temperature and Top_P (I talked here about what these parameters do).

The choice of models

Free text models

  • Venice Uncensored 1.1 (Dolphin-Mistral-24B-Venice-Edition), 32K context window. “Co-released by Venice and Dolphin, this model is built on Mistral’s 24B foundation. It offers an uncensored AI experience by removing artificial content restrictions, leading the field with a remarkably low 2.20% censorship refusal rate.” It didn’t impress me in any way!
  • Venice Medium (Mistral-Small-3.1-24B-Instruct-2503), 128K context window. “This 24-billion-parameter model provides advanced conversational and reasoning capabilities across dozens of languages. A key feature is its vision capability, allowing it to analyze images in addition to text.” Despite a larger context, it’s still a small model, and it shows. I got from it many robotic answers and answers that repeated large portions from previous answers!

Pro text models

  • Venice Reasoning (QwQ-32B), 32K context window. QwQ is based on Qwen2.5 by Alibaba Cloud. It’s quite old, as you can now use Qwen3. However, unlike Qwen3, which unified the thinking and non-thinking modes, QwQ‑32B is a direct specialization of Qwen 2.5‑32B, optimized for deep reasoning.
  • Venice Small (Qwen3-4B), 32K context window. “The latest 4-billion-parameter model in the Qwen series, featuring a unique “thinking mode” for complex reasoning (math, coding) and a “non-thinking mode” for efficient general dialogue. It offers enhanced reasoning, strong agent capabilities, and multilingual support for over 100 languages.” Yeah, it’s Qwen3, but the smallest of it, with a paltry 4B parameters! Why would anyone pay for such a tiny thing that could be run locally?
  • Venice Large (Qwen3-235B-A22B), 128K context window. The real thing, but even this newer one is terribly stupid (as tested upstream), not knowing what “A22B” means! As per ChatGPT: “Qwen3-235B-A22B is a Mixture of Experts (MoE) model with a total of 235 billion parameters, but only 22 billion are used during inference (per forward pass).” Not worth paying, unless, maybe, if it’s less censored than usual. But it doesn’t say anything about that.
  • Llama 3.1 405B (Llama-3.1-405B-Instruct), 64K context window. Great, it has 405 billion parameters. It’s not the lastest in the 3 series, but it’s the largest one, as 3.3 only has 70 billion parameters, and Llama 4 has a different architecture, with only 17B active parameters.

Free models retiring “on 06/21/2025” but still working on 07/07 — retired on 07/09

Note regarding the text models:

  • Currently, Venice can review documents with up to approximately 250,000 characters. If you receive an error after uploading a document, it likely exceeds Venice’s current threshold.

Free image models

  • HiDream (HiDream-I1-Dev): “An innovative open-source image generative foundation model with 17 billion parameters, designed for state-of-the-art image generation quality across photorealistic, cartoon, and artistic styles. It achieves high human preference alignment (HPS v2.1 score) and best-in-class prompt adherence.”
  • Venice SD35 (stable-diffusion-3.5-large): “Venice’s custom-configured version of Stable Diffusion 3.5. It brings significant improvements in photorealism, creating images with more realistic textures, lighting, and details. This model also boasts better prompt understanding, accurately following complex instructions and natural language prompts. It reduces common artifacts like distorted faces and unnatural textures, and offers an enhanced artistic range from hyperrealistic photography to stylized illustrations.” The link goes to the official model, so “custom-configured” must refer to the parameters one can set in the app’s settings and possibly to others set by Venice.
  • FLUX Standard (FLUX.1-dev): “an open-weight, guidance-distilled model offering the best image generation to date, featuring excellent prompt-following, visual quality, image detail and output diversity. FLUX produces stunning image quality and is able to perform well with complex workflows, making it an exciting model for creative AI applications. The quality can come with a trade off. FLUX can be a little slower than other models on Venice. You can remedy this by using a smaller number of steps (suggested 10) in the settings to experiment, and then increase as you refine the image.”
  • FLUX Custom 1.1 Uncensored (getphat-flux-reality-nsfw): “an uncensored, LoRA-modified version of FLUX, which removes much of the censorship that was inherent in the underlying model. While this enables NSFW image generation, note that the “Safe Venice” setting still blocks NSFW unless you turn this setting off. There is a LoRA Strength slider in the options for this model. The lowest setting means it should act nearly identically to the normal FLUX model, and the highest setting is the most uncensored (but sometimes causes quality issues).” This model is not in public access!
  • Pony Realism Uncensored (372465): “Pony is a model that generates highly detailed and realistic images, particularly well-suited for creating images of people, landscapes, and still-life compositions. It can produce images with intricate details and textures, making it a great choice for applications that require high-quality image generation.” This model is not in public access!

Pro image models

  • Lustify SDXL Uncensored (lustify-sdxl-nsfw-checkpoint): “a community fine-tuned SDXL model specializing in highly detailed figurative art with a focus on photorealism. It excels at rendering dynamic lighting, intricate textures, and stylized human forms while maintaining artistic coherence. The model is versatile, handling both natural language and Danbooru-style prompts, making it suitable for a range of subjects.” This model is not in public access!

Note regarding the image models:

  • Features like “Enhance” and “Upscale” consume significantly more image requests per use than standard image generations. For example, “Enhance 4x” uses 16 image requests. Using these features will deplete your daily limit faster.

The abysmally stupid image models

I never cared much about creating images via a chatbot because all image generators are utterly stupid, except for ChatGPT when it creates Ghibli-style stuff.

As proof that the models used here are creating poor-taste images, here you have the demo images posted on the HF model pages for HiDream, Venice SD35 (“Most Artistic”), and FLUX Standard (“Highest Quality”):

OK, the second one feels more profesional IMHO. (As I’ll soon see, this bears no relevance in practice.)

🤖

For lack of a better idea, I used the following prompt with 4 models:

Draw me (in landscape mode) Linus Torvalds in tears, embracing Tux the penguin.

And here’s what I’ve got from HiDream, Venice SD35, FLUX Standard, and Pony Realism:

No comment.