If you’ve spent any time in the corner of the internet where people build elaborate, sometimes messy, and often very "uncensored" stories with AI, you’ve hit the wall. You know the one. The "I can't answer that" wall. It’s why people flock to platforms like Chub.ai. But once you’re there, the settings menu looks like the cockpit of a 747. You’re staring at names like Mars, Mercury, Asha, and Mistral, wondering which one is actually going to write a coherent response and which one is just going to hallucinate gibberish.
The short answer is that Chub doesn't just use "a" model. It’s a multi-layered ecosystem.
Most people think Chub is just a website like Character.ai. It’s not. It’s a frontend that hooks into whatever brain you decide to give it. If you’re paying for a subscription, you’re using their internal "Chub Lore" models. If you’re a power user, you’re probably plugging in an external API key and using something else entirely.
The Internal Powerhouse: Mars and Mercury Explained
Chub’s bread and butter are its two subscription tiers: Mercury and Mars. Honestly, the naming convention is a bit confusing because these aren't the models themselves—they are the service levels that give you access to specific finetuned models.
Mercury: The Budget 13B Workhorse
Mercury is the $5-a-month entry point. For the price of a coffee, you get unlimited messages using models in the 13B (13 billion parameters) range. Currently, the star of the show here is MythoMax 13B.
MythoMax is basically a legendary "Frankenstein" model. It’s a merge of several Llama-2 finetunes designed specifically for roleplay. It’s surprisingly good at following instructions, but let’s be real—it’s 2026. A 13B model has its limits. It can get "loop-y" if the chat goes on too long, and it might lose the plot if your character description is 3,000 words of dense lore.
Mars: The Heavy Hitter
Then there’s Mars. This is the $20-a-month tier, and it’s where the "real" roleplay happens for most serious users. When you subscribe to Mars, you generally get access to two main flavors:
- Asha (70B): This is the flagship. It’s a massive 70-billion parameter model. It’s smarter, more creative, and significantly better at "reading between the lines" than the Mercury models. It handles complex group chats with multiple characters without forgetting who is standing where.
- Mixtral (8x7B): This uses a "Mixture of Experts" architecture. It’s fast. Very fast. It tends to be more logical than Asha but sometimes lacks that "flowery" prose that roleplayers crave.
Why "Uncensored" Actually Matters Here
You’ll see the word "uncensored" thrown around Chub a lot. In the AI world, this usually means the model has been finetuned on datasets that haven't been scrubbed of "edgy" or adult content.
Most mainstream models like GPT-4o or Claude 3.5 have heavy safety layers. If your story gets a little too violent or spicy, they’ll wag a finger at you. The models Chub hosts—Asha and Mercury—are specifically designed to ignore those guardrails. They follow the user’s lead. If you want a gritty noir story where people actually get hurt, these models won't lecture you on ethics.
The Trade-off of Logic vs. Freedom
There is a catch, though. Mainstream models from Google or Anthropic are billions of dollars smarter than open-source finetunes.
Users often find themselves in a tug-of-war. Do you use Google Gemini 3 Flash via an API key for its massive 1M+ context window and incredible logic, even if you have to "jailbreak" it to get past the filters? Or do you stick with Mars Asha because it just gets the vibe of a story immediately, even if it occasionally forgets what color shirt you're wearing?
Connecting External Brains: The API Side
Chub isn’t just a walled garden. One of its best features is the API Connection tab. This is where the true experts live. You can take the Chub interface—which has great lorebook support and character management—and plug it into basically any LLM on the planet.
- OpenRouter: This is the most popular choice. It acts as a middleman. You give OpenRouter some credits, and you can use Claude 3.5 Sonnet, Llama 3.1 405B, or even DeepSeek V3.
- NovelAI: Some people prefer the prose style of Kayra. You can plug your NovelAI key into Chub to use their storytelling-specific models.
- Self-Hosting: If you have a monster PC with a couple of RTX 4090s (or 5090s by now), you can run models locally using KoboldCPP and link it to Chub. This is the ultimate privacy move. Nothing leaves your house.
Context Windows: The Secret Sauce
If your character starts acting "dumb" after 50 messages, it’s usually a context window issue.
Most of the built-in Chub models like Asha operate with an 8K to 32K token context. That sounds like a lot, but a "token" is roughly 0.75 words. Between the character’s long backstory, the "Lorebook" entries (world-building notes), and the actual chat history, that space fills up fast.
When the context window is full, the AI has to start "forgetting" the earliest parts of the chat. This is why some users prefer using Mistral Small 24B or Gemini via API; they offer much larger windows (up to 96K or even 1M+), meaning the AI can remember a tiny detail from three days ago without you needing to remind it.
Setting Up for Success: Actionable Steps
If you're just starting out, don't just pick a model and pray. The "Model" is only half the battle. The "Preset" is the other half.
1. Tweak Your Temperature
If the AI is being boring and repetitive, your Temperature is probably too low (around 0.7). Bump it up to 1.1 or 1.2 to give it some "creative juice." If it starts speaking in riddles or broken English, turn it back down.
2. Use the Lorebook Properly
Don't cram everything into the character description. Use the Lorebook feature to define locations or secondary characters. This ensures the model only "thinks" about the tavern when you’re actually at the tavern, saving precious context space.
💡 You might also like: Atomic Radius Trends on Periodic Table: What Most People Get Wrong
3. Choose Based on Your Needs
- Casual, light RP? Mercury ($5) is fine.
- Deep, narrative-driven stories? Mars ($20) or a Claude API key via OpenRouter.
- Technical/Experimental? Use the free Mistral 24B options often found in the "Free" or "Mobile" tiers of the site.
The reality of 2026 is that the "best" model changes every month. Today it might be Asha; tomorrow it might be a new Llama-4 finetune. The beauty of Chub is that it doesn't lock you into one brain. You can swap the "model" like a cartridge in a video game console until you find the one that writes exactly the way you want it to.
To get the most out of your current setup, go into your Generation Settings and check your Repetition Penalty. If your bot is repeating your own words back to you, set this to about 1.1. It’s a small change, but it’s often the difference between a frustrating loop and a story that actually moves forward.