Best Free Model for Roleplay (2026): OpenRouter's Free List, Gemini's Free Tier and What Each Costs You
The best free model for roleplay on 2 October 2026 is NVIDIA’s Nemotron 3 Ultra, on OpenRouter’s free tier. It has a 1M-token context window, it costs nothing, and it is the free model that real roleplayers use the most: on Janitor AI it is the second most used model of any kind. Google’s Gemma 4 31B is the other free model those players run, and Gemini 3.8 Flash on Google’s own free API tier is the strongest free option outside OpenRouter.
“Free” always has a price that is not money. With free models you pay in daily limits, in privacy, in waiting, and in permanence: free models disappear with no warning. This page lists what is free today, shows which of those models roleplayers actually use, and is honest about each of those four costs.
Everything below was checked on 2 October 2026: OpenRouter’s model list, its documentation and its public usage pages for Janitor AI and SillyTavern, plus Google’s API pages and AI Horde’s live status.
A note on method. We have not run a controlled roleplay test of these models. Free lists change every few weeks, and a test would be out of date before it was finished. What we can report is what is listed, what the limits are, and what players have chosen.
If you want a free platform instead of a free model — a place to play with no setup at all — that is a different page: the best free AI roleplay.
The Short Answer
| You want… | Use | The catch |
|---|---|---|
| The free model roleplayers use most | Nemotron 3 Ultra on OpenRouter | 50 requests a day; thinks before it answers |
| A free model with faster replies | Gemma 4 31B on OpenRouter | 50 requests a day; smaller context window |
| A free model from a large lab | Gemini 3.8 Flash on Google’s free API tier | Your content is used to improve Google’s products |
| No setup at all | Janitor AI’s JLLM or AI Dungeon’s free models | Small context windows |
| No account, no key | AI Horde | Slow, and run by volunteers |
| No limits, ever | A local model on your own PC | You need the hardware |
| The best value overall | Buy 10 OpenRouter credits once | It is $10 |
Four Kinds of Free
People mean four different things by “free model”, and they behave differently.
- A free tier on an API. A company lets you use a model through its API at no charge, with limits. OpenRouter’s free models and Google’s free Gemini tier are this kind. You connect them to a frontend such as SillyTavern or Janitor AI.
- A volunteer network. AI Horde runs models on computers that volunteers lend.
- A free model built into a platform. Janitor AI and AI Dungeon both include models you can use with no key at all.
- A local model. You download a model and run it on your own graphics card. It costs electricity and hardware, and nothing else.
Most searches are about the first kind, so most of this page is too.
OpenRouter’s Free Models
OpenRouter puts hundreds of models behind one API key. Some of them have a free version, marked with :free at the end of the model name. On 2 October 2026 there were 17 models with a free version, out of 465 models in total.
Not all 17 are useful for roleplay. Three are coding models, one is a medical model, one is a content-safety filter, one is built to process video and audio for other programs, and one is a very small model with 2.6 billion parameters. That leaves ten general-purpose models. OpenRouter also listed two more general models at a price of zero without the :free mark.
Here are those twelve.
| Model | Model name to type | Context window | Thinking |
|---|---|---|---|
| Nemotron 3 Ultra (NVIDIA) | nvidia/nemotron-3-ultra-550b-a55b:free | 1M | On by default |
| Gemma 4 31B (Google) | google/gemma-4-31b-it:free | 262K | Optional |
| Nemotron 3 Super (NVIDIA) | nvidia/nemotron-3-super-120b-a12b:free | 262K | On by default |
| Nemotron 3.5 Lightning (NVIDIA) | nvidia/nemotron-3.5-lightning:free | 1M | Optional |
| Gemma 4 26B A4B (Google) | google/gemma-4-26b-a4b-it:free | 262K | Optional |
| Qwen3.8 27B | qwen/qwen3.8-27b:free | 262K | On by default |
| Inkling (Thinking Machines) | thinkingmachines/inkling:free | 1M | On by default |
| Inkling Small (Thinking Machines) | thinkingmachines/inkling-small:free | 1M | On by default |
| Dots3-Note Preview (Dots Studio) | dots-studio/dots-3-note-preview:free | 512K | Optional |
| Apodex 1.1 Mini | apodex/apodex-1.1-mini:free | 262K | Optional |
| Ling 3.1 Flash (inclusionAI) | inclusionai/ling-3.1-flash | 262K | On by default |
| Space Bunny Alpha (anonymous) | stealth/space-bunny-alpha | 1M | Always on |
Three notes on this table.
Some entries have an end date. OpenRouter’s list shows an expiry date on a few free models. Space Bunny Alpha is listed to end on 5 October 2026 and Dots3-Note Preview on 31 December 2026. A model with an end date is a trial, not a home for a campaign.
“Anonymous” means what it says. Space Bunny Alpha is listed by a provider called “stealth”, and its description calls it “an anonymous large model”. Models like this are usually offered free so that the maker can collect feedback before a launch. Do not put anything in it that you would mind a stranger reading.
There is no free Gemini, Claude, GPT, DeepSeek or Mistral model on this list. Google’s free models here are from its open Gemma family. For free Gemini, you go to Google directly; see below.
OpenRouter also has a model name that picks for you: openrouter/free. Its documentation says it “automatically selects a free model at random from the available free models”. For roleplay that is a poor idea, because your character’s voice changes with the model, and it could change on every message.
Which Free Models Roleplayers Actually Use
A list of twelve does not tell you which one to pick. Usage does.
OpenRouter publishes the models that each connected app uses the most. We read the pages for the two biggest roleplay frontends on 2 October 2026. In the last 30 days:
| App | Free-listed model | Rank among all models in that app | Tokens |
|---|---|---|---|
| Janitor AI | Nemotron 3 Ultra | #2 | 96B |
| Janitor AI | Nemotron 3 Super | #19 | 12.3B |
| SillyTavern | Nemotron 3 Ultra | #15 | 7.77B |
| SillyTavern | Gemma 4 31B | #17 | 7.43B |
One caution. The usage pages show a model’s name, not whether people ran its free or its paid version. All three of these models have both. We think the free version explains most of this use, because these models rank far higher in roleplay apps than their paid prices alone would suggest. That is our reading, not a published fact.
What the table does show clearly:
- Nemotron 3 Ultra is the free model roleplayers have settled on. On Janitor AI only one model, DeepSeek V3.2, is used more.
- Gemma 4 31B is the second choice, and it appears among SillyTavern users, who tend to be more experienced.
- No other free-listed model appears in either top 20.
That is a strong signal. When thousands of players with no budget all end up on the same two models, those are the two to try first. Full lists for each app are in the best model for Janitor AI and the best model for SillyTavern.
Nemotron 3 Ultra or Gemma 4 31B?
| Nemotron 3 Ultra | Gemma 4 31B | |
|---|---|---|
| Maker | NVIDIA | |
| Size | 550B parameters in total, 55B active at a time | About 31B |
| Context window, free version | 1M tokens | 262K tokens |
| Thinking | On by default | Optional |
| Longest reply | 65,536 tokens | 32,768 tokens |
| Can you run it at home? | No | Yes, on a 24GB graphics card |
Pick Nemotron 3 Ultra for long campaigns: the context window is four times larger, and it is the bigger model. Its description on OpenRouter calls it a “frontier-reasoning and orchestration model”, which tells you it was built for reasoning tasks, not fiction. Roleplayers use it anyway, in large numbers.
Pick Gemma 4 31B if replies on Nemotron start too slowly for you. Thinking is optional on Gemma, so it can answer at once. It is also a model you can later run on your own PC, which means that what you learn about prompting it stays useful if you leave the free tier.
The Limits: What 50 Requests a Day Means
OpenRouter’s documentation sets out the free limits in one small table.
| Credits you have bought, in total | Requests per minute | Requests per day |
|---|---|---|
| Fewer than 10 | 20 | 50 |
| At least 10 | 20 | 1,000 |
Five things follow from it.
1. The limit counts requests, not tokens. This is the opposite of paid use. On a paid model a long context costs you money on every turn. On a free model, a turn with 100,000 tokens of history counts the same as a turn with 2,000. So on the free tier, you can use a large context without any penalty except speed.
2. Every swipe is a request. If you reroll each reply twice, 50 requests is about 17 turns of story. That is one short session.
3. Other things can use requests too. In a frontend such as SillyTavern, an automatic summary of your chat is also a request to the model. If your daily count disappears faster than you expect, look for features that call the model in the background.
4. A negative balance blocks free models. The documentation says: “If your account has a negative credit balance, you may see errors, including for free models.”
5. New accounts or keys do not help. OpenRouter says that making additional accounts or API keys “will not affect your rate limits”.
The $10 that changes everything
Look at the second row of that table. Once you have bought 10 credits in total, once, your free limit goes from 50 to 1,000 requests a day. The documentation is specific that the tier depends on credits purchased, all time, not on your current balance.
So the best free setup is not quite free. It costs $10, one time. And you still have the credits. The same $10 buys about 790 long turns on DeepSeek V4.1 Flash, or thousands on the cheapest models in the table further down. For anyone who plays more than a couple of times a week, this is the real answer to the question this page is about.
What Free Costs in Privacy
Free models are often free because someone wants the data.
On OpenRouter, the documentation explains that each provider has its own policy, and that in your account settings “you can set whether you would like to allow routing to providers that may train on your data”. It adds: “There are separate settings for paid and free models.” In practice, many free models only work if you allow that for free models. Janitor AI’s own help centre tells its users to switch that setting on to use free models.
On Google’s free API tier, Google says the content you send is used to improve its products. On the paid tier it is not.
On AI Horde, a volunteer’s computer processes your chat. SillyTavern’s documentation warns that a malicious worker could log your prompts and replies, and says: “When using Horde never send any personal information such as names, email addresses, etc.”
None of this is hidden, and none of it is a scandal. It is the trade. If you would be unhappy for a stranger to read a chat, do not send it through a free model.
Free Models Disappear
The fourth cost is that nothing free is promised to stay.
The clearest example is DeepSeek. Through most of 2025, free DeepSeek on OpenRouter was the free roleplay model. Guides, presets and even Janitor AI’s own quickstart article still give the model name deepseek/deepseek-chat-v3-0324:free. On 2 October 2026 none of OpenRouter’s free models is a DeepSeek model. The community-made “Chimera” blends that were free for a while are not on the list at all.
Chutes is the other example. It was once the best-known source of free DeepSeek, and it ended its free access in steps between August 2025 and March 2026. Our Chutes AI guide has the dates.
Nemotron 3 Ultra’s free version was listed on 4 June 2026. It has lasted four months. Nobody has promised a fifth.
Two habits protect you:
- Check the list before you follow any guide, including this one. The names in the table above were right on 2 October 2026.
- Do not tune a long campaign around one free model’s style. When it goes, the next one will write differently.
Gemini’s Free Tier
If you want a free model from one of the large labs, Google is the one that offers it.
Through the API. Google’s API has a free tier for its Flash models, including the current Gemini 3.8 Flash. You get a key from Google AI Studio and connect it to a frontend. SillyTavern’s documentation lists Google AI Studio as having “a free tier with rate limits (Gemini Flash)”. Google does not publish the free-tier limits on its pricing page; you see them in AI Studio. Gemini 3.1 Pro has no free API tier.
Through the Gemini app. The app’s free plan gives you a context window of 32,000 tokens. That is enough for a short adventure and too small for a long campaign. Paid plans raise it.
The cost, again, is privacy: free-tier content is used to improve Google’s products.
Which Gemini model, and how the free and paid tiers compare, is in the best Gemini model for roleplay. For playing an RPG in the Gemini app itself, see the Gemini RPG guide.
What about Mistral and Qwen? People search for both. OpenRouter has one free Qwen model today, Qwen3.8 27B, in the table above. It has no free Mistral model. SillyTavern’s documentation says Mistral’s own API has a “Free Tier with rate limits”; we did not verify its current terms, so we do not describe them here.
Free Without a Key: Built-In Models and AI Horde
Janitor AI’s JLLM. Janitor’s own model is free, with no message limit and no setup. The free plan gives it a context of around 9,000 tokens. See Janitor AI pricing.
AI Dungeon’s free models. AI Dungeon’s free tier has nine story models and no limit on play, with up to 4,000 tokens of context. See AI Dungeon membership tiers.
AI Horde. On 2 October 2026 its live list showed 29 text models, most of them community roleplay fine-tunes. It needs no account, and SillyTavern connects to it out of the box. Guests are served at low priority, so replies can take a long time. We cover how to choose a Horde model in the best model for SillyTavern.
The chat apps. The Gemini app, Claude and DeepSeek’s chat app all have free plans. They work for a quick scene with a pasted prompt. They are general assistants, not roleplay frontends, so there are no character cards and no lorebook. Our RPG prompts page has setups you can paste into any of them.
Free Forever: A Local Model
A model on your own PC has no daily limit, no privacy trade and no end date. It costs a graphics card. With about 8GB of VRAM you can run an 8B model; with 24GB you can run Gemma 4 31B, the same model as on the free list.
That is why Gemma is a useful free model to learn on. If you like it on OpenRouter’s free tier, you know what a 24GB card would give you with no limits at all. Local models for AI RPGs explains what it takes.
Nearly Free: What a Few Cents Buys
Before you accept the limits of free, look at what the cheapest paid models cost. This is one long session: 200 turns, each sending 40,000 tokens of context and getting a 500-token reply. It is the same session we price on all our model pages.
| Model | 200 turns | Context window |
|---|---|---|
| DeepSeek V4 Flash, April version | $0.23 | 1M |
| Gemma 4 31B, paid version | $0.75 | 262K |
| GLM 5.3 Flash | $1.25 | 1M |
| DeepSeek V3.2 | $2.28 | 164K |
| DeepSeek V4.1 Flash | $2.52 | 1M |
These are our arithmetic on OpenRouter’s listed prices, before any caching discount and without thinking. With a smaller context they fall further: at 16,000 tokens every figure is less than half.
Paid use has no daily cap, and you can choose providers that do not train on your prompts. For a few cents an evening, three of the four costs of free go away. Which DeepSeek model to choose is in the best DeepSeek model for roleplay, and how OpenRouter billing works is in OpenRouter for roleplay.
Getting the Most From a Free Model
- Turn thinking off if the model allows it. Thinking makes replies start late. On a free model you are not paying for it, but you are waiting for it.
- Use the context window. The free limit counts requests, not tokens, so a large context costs you nothing.
- Swipe less. Each reroll is one of your 50. Edit a weak reply by hand instead of rerolling it.
- Switch off background features that call the model, such as automatic summaries, or they will spend your requests for you.
- Keep your setup portable. Put what matters in the character card, the lorebook and your own notes, not in tricks that only work on one model. Our guide to why AI campaigns fall apart explains what to keep and where.
Our Recommendation
- Try Nemotron 3 Ultra first, then Gemma 4 31B if you want faster replies.
- If you play more than a little, buy 10 OpenRouter credits once. It turns 50 requests a day into 1,000.
- For a large-lab model at no cost, use Gemini 3.8 Flash on Google’s free API tier, and accept that your content is used.
- For anything private, do not use a free model. Pay the few cents, or run a local model.
- Check the list today. If the model name in a guide is no longer there, that is why your setup fails.
Frequently Asked Questions
What is the best free model for roleplay?
On 2 October 2026, NVIDIA’s Nemotron 3 Ultra on OpenRouter’s free tier. It has a 1M-token context window and it is the free model that roleplayers on Janitor AI and SillyTavern run the most. Google’s Gemma 4 31B is the second free model those players use, and Gemini 3.8 Flash on Google’s own free API tier is the strongest free option outside OpenRouter. Free lists change often, so check the model is still listed before you set it up.
What is the best free OpenRouter model for roleplay?
Nemotron 3 Ultra, with the model name nvidia/nemotron-3-ultra-550b-a55b:free. It is the second most used model of any kind among Janitor AI users on OpenRouter. If you want replies that start faster, Gemma 4 31B, google/gemma-4-31b-it:free, lets you turn thinking off. OpenRouter listed 17 models with a free version on 2 October 2026, and about ten of them are general-purpose models you could roleplay with.
What are the limits on OpenRouter’s free models?
20 requests a minute for everyone. 50 requests a day if you have bought fewer than 10 credits in total, and 1,000 requests a day once you have bought at least 10 credits. The limit counts requests, not tokens, and every swipe or reroll is a request. If your account balance is negative, free models stop working too.
Is there a free DeepSeek model for roleplay?
Not on OpenRouter. On 2 October 2026 none of its free models was a DeepSeek model, although older guides still give the name deepseek/deepseek-chat-v3-0324:free. DeepSeek’s own chat app is free, but it is a general assistant, not a roleplay frontend. The cheapest popular DeepSeek model on OpenRouter costs about $0.23 for a 200-turn session, so nearly free is still available.
Is Gemini free for roleplay?
Yes, in two ways. Google’s API has a free tier for its Flash models, including Gemini 3.8 Flash, which you can connect to SillyTavern or Janitor AI. And the Gemini app has a free plan with a 32,000-token context window. On the free API tier, Google says the content you send is used to improve its products. Gemini 3.1 Pro has no free API tier.
Are free models private?
Less than paid ones. OpenRouter has a separate account setting for whether free models may be routed to providers that train on your prompts, and many free models only work with it switched on. Google says free-tier API content is used to improve its products. AI Horde runs on volunteers’ machines and its documentation tells you never to send personal information. If privacy matters, a paid model at a few cents is the better choice.
Why did my free model stop working?
There are four common reasons. You reached the daily limit of 50 or 1,000 requests. The model was removed from the free list, which happens often. Your OpenRouter balance is negative, which blocks free models too. Or your privacy setting does not allow the provider that hosts that free model.
Is it worth paying a little instead of using a free model?
Usually yes. Buying 10 credits on OpenRouter once raises the free limit from 50 to 1,000 requests a day, and the credits stay yours to spend. The same $10 pays for about 790 long turns on DeepSeek V4.1 Flash, or thousands on the cheapest models. For most players the honest best free setup is to pay $10 once.
The free models, their names, context windows, thinking defaults, reply limits, listing dates and end dates are from OpenRouter’s model list on 2 October 2026. The request limits, the negative-balance rule, the free router and the training settings are from OpenRouter’s documentation. The usage figures are from OpenRouter’s public app pages for Janitor AI and SillyTavern (last 30 days) on the same day. Gemini’s free tier and the Gemini app’s context window are from Google’s own pages, checked on 1 October 2026. AI Horde’s model count is from its public status API. The sorting of free models into roleplay-suitable and not, the reading that free versions drive the usage ranks, and every session cost are our own. We have not run a controlled comparison of how these models write.
Product names and trademarks are the property of their respective owners. Arcanum is not affiliated with OpenRouter, NVIDIA, Google, AI Horde or any model maker named here.