Free OpenRouter Models for Roleplay: November 2026

Checked October 30, 2026: OpenRouter lists 16 free (:free) models. None was built for roleplay. Three are coding or moderation specialists, two more coding models go away on October 31, and two only answer inside agentic harnesses. That leaves nine general models you can put in Janitor AI or SillyTavern, and in our requests only four of them answered every time.

Start with Dots3-Note Preview (dots-studio/dots-3-note-preview:free). It answered 6 of 6 of our requests, has a 512,000-token context and 460,800-token output limit, and its free endpoint doesn't train on your prompts. OpenRouter lists it until December 31, 2026. If you don't mind your chats being used for training, the three free NVIDIA Nemotron models also answered 6 of 6 and have 1M-token context.

The free DeepSeek ID in Janitor AI's own setup guide no longer has a free endpoint. deepseek/deepseek-chat-v3-0324:free returns zero endpoints on OpenRouter today. Replace it with one of the IDs below.

We measured availability, not writing quality. For the full free catalog, see our OpenRouter free models guide.

The free models you can use in a chat frontend

As of 2026-10-30 from OpenRouter's models API, each free model's /endpoints API, the data policy on each model's OpenRouter page, and our own requests.

How we computed it: we kept :free IDs with zero input and output prices and text output, then removed the models listed in the next section. "Answered" counts single-turn requests from our free-tier key that got a reply on the first try: two a day on October 28 (16:16 UTC), 29 and 30 (12:34 UTC), n = 6 per model. Every miss on Gemma 4 and Xing4.0 was a 429 rate limit, not a wrong answer. Context, max output and uptime (uptime_last_1d) are the free endpoint's own values. Paid prices are the paid listing in the models API.

What the table says:

• The free models that respect your privacy were the hardest to reach. Gemma 4 and Xing4.0 don't train on prompts, and Xing4.0 is the only free endpoint that doesn't keep them at all. All three answered none of our six requests each, because their shared free pool was full. A 100% uptime figure means the provider was up, not that you'll get a turn.
• The models that always answered may train on your chats. All four NVIDIA endpoints in the table may use prompts for training, and their policy lets OpenRouter train on them too. If you've turned off training on free models in your OpenRouter privacy settings, OpenRouter won't route to them, and they'll fail for you.
• Dots3-Note Preview is the only model in both groups. It answered 6 of 6 and doesn't train on prompts. It has no paid version on OpenRouter, so plan a replacement before December 31.
• LFM2.5-2.6B is too small for long chats. Its free endpoint caps context at 65,536 tokens and replies at 8,192, and it answered 2 of 6.

The free models to skip for roleplay

North Mini Code answered 6 of 6 and doesn't train on prompts, so you can try it. Just know Cohere built it for code, not fiction. For what the coding models are good at, see our free OpenRouter models for coding.

Settings that matter for long chats

Sampler sliders. SillyTavern sends temperature, top P, top K, min P and repetition penalty. Each free endpoint lists the parameters it supports, and most list far fewer:

If your preset relies on repetition penalty or min P to keep a long story from looping, only Xing4.0 and LFM2.5-2.6B list them. On the other models, don't count on those sliders doing anything.

Thinking. OpenRouter lists reasoning as on by default for Xing4.0, Nemotron 3 Ultra and Nemotron 3 Nano Omni, and as off by default for both Gemma 4 models. Thinking spends output tokens and time before the reply starts. If replies are slow or cut off, raise the max response length first.

Context. A 1M-token window doesn't make a model remember a 1M-token chat well; we didn't test recall. Check after a few hundred messages whether the character still uses facts from early in the story.

Moderation. OpenRouter marks none of these 16 endpoints as moderated (is_moderated: false). What a model will and won't write still comes from its own training, which we didn't test.

Janitor AI setup with a free OpenRouter model

From Janitor AI's proxy quickstart:

1. Create an OpenRouter account and a key under Keys.
2. In a Janitor AI chat, open the gear icon, then API Settings, then Add Configuration.
3. Model Name: paste the exact ID, including :free, for example dots-studio/dots-3-note-preview:free.
4. Proxy URL: https://openrouter.ai/api/v1/chat/completions
5. API Key: your OpenRouter key, with no extra spaces.
6. Click Add, then Save Settings, reload the page and send a message.

The quickstart's example model, deepseek/deepseek-chat-v3-0324:free, has no free endpoint as of October 30, 2026, so don't copy it.

SillyTavern setup

From SillyTavern's OpenRouter docs: set the API to Chat Completion, choose OpenRouter as the source, click Authorize or paste a key, click Connect, pick a model ending in :free, and use Test Message to check it works.

Free limits, and what paid costs per message

OpenRouter's limits documentation sets these limits for :free models:

Every reply is a request, and so is every swipe or regeneration. Fifty a day goes fast in a roleplay session. A negative credit balance can also block free models.

If you outgrow the free tier, here's what a paid version of the same model costs. We assume a reply that reads 8,000 tokens of chat history and character card and writes 400 tokens:

How we computed it: 1,000 × (8,000 × input price + 400 × output price) ÷ 1,000,000, using the paid listing in the models API on October 30, 2026. Longer histories cost more, because the whole history is sent with every reply. Paid endpoints have their own data policies, so check them before you switch.

Check your free IDs against the live list before each session:

curl -s https://openrouter.ai/api/v1/models | jq -r '.data[]
  | select(.id | endswith(":free"))
  | [.id, .context_length, (.expiration_date // "-")] | @tsv'

Writing stories for work in Teamday

Teamday is built for business work, not companion chat. If you write fiction-style content for your company, such as campaign scripts, newsletters or product stories, Maya, Teamday's content creator, drafts it from your brief for you to review. The free OpenRouter endpoints in this guide are not in Teamday's model picker. See current pricing.

Related guides

• OpenRouter free models: the full catalog
• Best free OpenRouter models for coding
• OpenRouter top models: usage rankings and prices