How “free” actually works on Janitor
Every Janitor chat runs on some model. Free means one of two things: the built-in JanitorLLM, which costs nothing and needs no setup, or an outside model whose provider offers a free tier that you connect through a proxy. The second category is where the quality lives, and it is also where the ground moves fastest.
Two rules before the list. Free quotas do not roll over: “No, these don’t stack. If you only used 50 yesterday, you don’t get 350 today,” as the community megathread puts it. And free routes die without notice: Chutes went paid in 2026, and the free DeepSeek variants left OpenRouter. Every “best free models” list older than a few months, including the biggest Reddit guides, now recommends models you cannot actually run for free. I verified every availability claim below against live sources on July 26, 2026.
The models, ranked for roleplay
The rankings come from the community’s own head-to-head, a 514-upvote overview by an active roleplayer, plus the megathread’s model reviews; the quoted vibes are theirs. The availability status on each entry is mine, checked July 26, 2026. Where a model is no longer free anywhere, it stays on the list, because knowing where it went is the information everyone else’s stale lists are missing.
DeepSeek V3 0324: the community’s daily driver
The all-rounder the threads keep coming back to: strong prose, handles “passionate scenes and authentic, foul-mouthed dialogue without getting preachy,” in the overview’s words, with the recurring caveat that it “tends to make characters soft.”
Where to run it now: not free anywhere I could verify. Gone from OpenRouter’s free tier; the cheap route is DeepSeek’s official API with a small top-up, and the community’s verdict on the free workarounds is blunt: “they never really work. If you’re gonna use DeepSeek, just pay.”
DeepSeek R1 0528: “The Witty Humorist & Canon Lawyer”
The reasoning variant, ranked by the community for humor and canon accuracy. Reasoning models need one Janitor-specific setting: max new tokens at 0, or you get thinking text instead of replies.
Where to run it now: same status as V3, official API only. If you connect it, the proxy guide’s error table covers the “No response from bot” quirk it is famous for.
Chimera R1T2: the humanlike specialist
The community’s pick for slice-of-life and sitcom-energy roleplay, “The Creative & ‘Humanlike’ Specialist,” with a documented weakness: “can struggle with memory in very, very long chats.”
Where to run it now: its free OpenRouter variant is gone. Availability now depends on paid hosting; check the live model list before planning around it.
Qwen3 235B: “the model for control freaks”
The overview’s pick for users who want obedience over flair: stable, logical, follows instructions, least creative of the favorites. Community NSFW reports are mixed and preset-dependent.
Where to run it now: the free variant the community used has left OpenRouter; paid hosting remains.
Gemini 2.5 Flash and Pro: the biggest free quotas left
The one first-tier family still genuinely free: Google AI Studio issues free Gemini keys with per-model daily caps. The community ranks Pro as “The Master Storyteller & World-Builder,” the undisputed king of prose, with two costs: it is “VERY promt depended,” and its filter behavior on explicit content varies, which the threads manage with custom prompts. Flash is the volume pick; older community write-ups put its daily allowance well above Pro’s, and Google now keeps the exact caps on its live rate-limit pages, so check there rather than trusting any fixed number.
Where to run it now: free at Google AI Studio, the last big free quota standing.
Mistral NeMo: the one to skip
Included because the guides keep recommending it and the community keeps warning against it: “NeMo tends to ignore prompt instructions and relentlessly speaks for the user, worse than Deepseek.” A negative consensus is information too; spend your setup time on the entries above.
Free quotas compared: what is actually left
| Route | Status, Jul 26, 2026 | Cap reality |
|---|---|---|
| JLLM built-in | Free, no setup | The baseline; no key, no quota to manage |
| Google AI Studio (Gemini) | Free tier live | Per-model daily caps, exact numbers on Google’s live page |
| OpenRouter free list | Live, but transformed | 15 models, all corporate-aligned; the roleplay favorites are gone |
| DeepSeek official API | Paid | Small top-up, community-reported ~$2 minimum, lasts a long time at roleplay volumes |
| Chutes | Paid since 2026 | The cautionary tale: “Free things rarely stay free” |
How to connect a model to Janitor AI
Short version: provider account, API key, then Janitor’s API Settings with the provider’s URL, the exact model string, and the key, pasted not typed, saved and hard-refreshed. The full walkthrough, current-UI caveats included, plus the fix for every error code you will meet on a free tier, is in my Janitor AI proxy guide.
When the rate limits hit
A free-tier evening ends the same way often enough that the community has a catchphrase for it: 429. Daily caps count blocked generations on some providers, which produced the megathread’s saddest quote: “my 50 messages goes to waste by this error.” Multi-account workarounds get proposed in every thread and they run against provider terms, which ban duplicate accounts; build your plans on the caps, not around them.
If you just want uncensored chat that works out of the box
The other honest answer to the rate-limit evening: some of us eventually stop wanting a hobby and start wanting a product. These are the apps I paid for and tested that deliver uncensored roleplay with zero model management.