Which LLM Powers Which AI Companion App
A maintained lookup table of what is actually running underneath every major companion platform, what each company discloses, and what stays undisclosed. Checked and dated.
By Ash Kepler · Jul 25, 2026 · 6 min read
Your companion is a character card sitting on top of a language model. The card is yours. The model belongs to the company, and most of them will not tell you what it is.
This table is what is publicly known as of the date above — model-provenance table v1.0, verified 25 Jul 2026 and rechecked quarterly. Where a platform discloses nothing, the row says Undisclosed rather than guessing.
The table
| Platform | Model underneath | Disclosed? | Can you change it? | |---|---|---|---| | Janitor AI | JLLM (proprietary, beta late 2023) | Named, not documented | Yes, via bring-your-own-key | | Character.AI | Proprietary in-house | Named only at version level | No | | Replika | Proprietary, rebuilt as Replika 2.0 | No | No | | Nomi | Proprietary, vector memory | No | No | | Kindroid | Proprietary, vector memory | No | No | | SillyTavern | None. You supply the backend | N/A | Entirely | | Chub AI | API workflow, your choice | N/A | Yes | | Candy AI | Undisclosed | No | No | | CrushOn | Undisclosed | No | No | | SpicyChat | Undisclosed | No | No |
Three groups fall out of that, and they behave completely differently.
Group one: proprietary and closed
Character.AI, Replika, Nomi, and Kindroid all run models they built and do not describe. You get a product, not a component.
Janitor sits at the edge of this group with JLLM, its proprietary model launched into beta in late 2023 and designed specifically for unfiltered roleplay, without the safety layers found in commercial models. The platform ran on OpenAI's GPT-3.5 Turbo and GPT-4 APIs during its 2023 surge, and moved to JLLM partly because OpenAI's usage policy on adult content was producing account bans.
Two of these use an architecture worth knowing about. Kindroid and Nomi are the platforms currently using vector memory, where key details are stored in a database and retrieved when relevant, rather than relying purely on the context window. That is why they benchmark better on long-term recall, and it is a genuine architectural difference rather than a marketing claim. Our guide to why your companion forgets covers what the alternative looks like.
Group two: you pick
Janitor's bring-your-own-key setup means you choose which model powers your characters, most commonly through OpenRouter. SillyTavern goes further and has no model at all, connecting to local and cloud backends with character cards, world info, group chats, and deep prompt control.
Current community favorites for that route, as of this update: GLM 5.2 and DeepSeek V4 for companion and roleplay backends, with Gemma 4 31B as the budget local option. One notable avoidance: Claude Fable 5 drew censorship backlash from the roleplay community and many roleplayers skip it despite the prose quality.
Our SillyTavern setup guide covers the install, and Janitor AI is the low-effort entry point to the same idea.
Group three: undisclosed
Candy AI, CrushOn, and SpicyChat publish nothing about what runs underneath. That is normal for the subscription NSFW tier and it is not evidence of anything sinister. It does mean you are buying an experience with no visibility into the component that produces it, and no notice when it changes.
Our Candy AI review and SpicyChat review cover what each delivers in practice, which is the only measurement available when the spec sheet is blank.
Why this table matters more than it looks
Three practical consequences.
Your character's behavior is downstream of a component you do not control. A platform can swap models on a Tuesday and your companion sounds different on Wednesday, which is the mechanism behind our guide to why your companion suddenly turned formal.
Undisclosed means unannounced. Platforms that never told you what they were running have no reason to tell you when it changes. The change log is the only way to track it, and we maintain one at the AI companion model change log.
Model choice is the only real hedge. Bring-your-own-key and self-hosted setups are the only configurations where nobody can change your companion's brain without your involvement. That is the actual argument for the harder path, and it is a better argument than the usual one about filters.
Methodology and corrections
Entries are marked disclosed only where the company names the model itself. Community consensus, benchmark inference, and third-party speculation are not treated as disclosure, which is why several rows read undisclosed even though guesses circulate.
This table is versioned and dated at the top of the page. If you have a documented correction, particularly a company statement we have missed, send it and it gets checked into the next revision.
If you want the surrounding decision rather than the component list, our roundup of the best AI companion apps sorts the same platforms by what they are actually like to use, and Candy AI, CrushOn, and SpicyChat are all covered there in detail.
questions