Best LLMs for Roleplay in 2026 (So far)
There is no universal best model for roleplay. A great RP model has to keep a character's voice, follow the scene, remember what matters, and write in a way you want to read for more than one reply. Those strengths do not always arrive together.
Our creative-model research gives a useful starting list. In the survey, GLM-5.1 was the most frequently named favorite, followed by GLM-4.7 and Kimi K2 Thinking. The same report also measured a separate score index where Kimi K2.6, GLM-5.1, GLM-5, and DeepSeek V4 Pro led the listed models. Treat those results as candidates to try, not a promise that one will suit every character. Survey results reflect what respondents had used, and availability, price, and model versions change.
What makes an LLM good for RP?
For most roleplay, instruction following and prose matter more than raw puzzle-solving. The research participants repeatedly asked for models that stay in character, avoid writing the user's actions, move the plot forward, retain story details, and match the genre. A model can be strong at general questions yet make a flat scene if it repeats your wording, resolves every conflict too easily, or turns every character into the same agreeable voice.
Best starting models by RP style
| Start with | Use it for | Why it is on the list |
|---|---|---|
| GLM-5.1 | A broad first test for character chat and prose-heavy scenes. | It was the clear favorite in the creative-model survey and also placed near the top of the separate score index. |
| Kimi K2.6 | Long-form scenes where you want to compare top-ranked writing output. | It led the research report's listed score index, so it is a sensible benchmark against your usual model. |
| GLM-5 or DeepSeek V4 Pro | Trying alternatives after GLM-5.1, especially when a prompt needs a different voice. | Both ranked close to the top of the score index. Run the same scene through each before deciding. |
| GLM-4.7 or Kimi K2 Thinking | Starting from strong community preference and testing how a model handles your instructions. | They were among the most often named favorites in the survey, despite lower index scores than some newer options. |
| Gemma 4 31B | A model to evaluate when you value WriterBench's strongest published result in this dataset. | It had the highest WriterBench total in the report's underlying benchmark data, but was named less often by survey respondents. |
The mismatch between the favorites and the benchmark is the point. A benchmark can measure useful writing traits, but it cannot know whether a model understands your vampire court, your goofy group chat, or the amount of initiative you want from an NPC.
How to test a roleplay model fairly
- Use one familiar character card, system prompt, lorebook, opening message, and generation setup.
- Give every candidate the same three-scene test: dialogue, a plot decision, and a scene that calls back to old story information.
- Read several turns. A strong opener does not prove the model will keep its voice or memory through a session.
- Check for the failures that matter to you: speaking for you, repeated phrasing, missed instructions, positivity bias, or formatting drift.
- Compare cost and speed only after the writing clears your bar. The best affordable model is the one you will keep using without dreading a regeneration.
Prompts and memory still matter
Most respondents used a prompt, a preset, or a lorebook alongside the model. That tracks with normal RP use: the model supplies the performance, while your setup supplies the cast, rules, and tone. Keep character facts concise, put non-negotiable formatting rules near the top, and use a lorebook for details that should return at the right moment. A clean setup makes a model comparison far more revealing.
On Composite, you can keep reusable prompts and lorebooks with an API key, then try candidates without rebuilding the whole setup each time. For a longer campaign, pick the context-management option that fits your account and test it early, before the story has months of history.
Our recommendation
Start with GLM-5.1 if you want the survey's most popular choice. Put it beside Kimi K2.6 and one alternate such as GLM-5 or DeepSeek V4 Pro, then test a scene you know well. If your priorities are character voice and story momentum, let those results outweigh a generic leaderboard. The right roleplay model is the one that makes your particular cast feel alive without taking control of the story from you.
Read the full creative-model research for the survey context, WriterBench methodology, and detailed results.