Model guide · 2026

The best AI models for roleplay, compared side by side.

LoreWeaver lets you choose the AI model that plays your story. This guide covers all 42, from what each one does well in roleplay and long stories to where it falls short. Every profile also gives its speed and the plans that include it.

Models
42
Makers
17
Checked
October 18, 2026

Written by the LoreWeaver team. Speeds are measured on LoreWeaver’s own replies, and every profile lists drawbacks.

The short answer

What is the best AI model for roleplay?

There isn’t one best model for every story, but a few stand out on LoreWeaver. GLM-5.3 Flash is the strongest all-rounder from the Plus plan up, and Muse Spark 1.3 holds up best over long campaigns on that plan. Kimi K3 and GLM-5.3 rank highest of LoreWeaver’s models on EQ-Bench’s creative writing board, while Aion 3.5 suits dark, high-tension drama. Mistral Nemo 12B is free and unlimited on every plan.

Updated

models to choose from, each with its own profile
42
AI labs behind them, from Anthropic to Xiaomi
17
the most story a model keeps in view, on Legendary
100,000 tokens
models that usually finish a reply in under 5 seconds
5

Every model

Find the model for your story. Then compare a few side by side.

42 models

Popular comparisons

Claude

by Anthropic · 1 model

Claude Haiku 5.5

Anthropic · Released October 2026

Best for quick, clean scenes

Anthropic’s Claude Haiku 5.5 came out on October 7, 2026, so no fiction or roleplay test covers it yet. On LoreWeaver it is one of the quickest models and keeps up to 100K of story in view on Legendary. Anthropic’s own testing found it quicker to refuse and moralize than other recent Claude models.

Plans
Plus and up
Speed
Very fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • SpeedOne of the quickest models on LoreWeaver, usually under 4 seconds a reply.
  • Long storiesKeeps up to 100K of story in view on Legendary.
Drawbacks
  • Dark themesAnthropic found it refuses and moralizes more than other recent Claude models, so dark scenes can meet resistance.
  • NarrationWrites best when it thinks at length, and LoreWeaver keeps its thinking short so replies stay quick.
  • Following directionThe Creative and Chaos presets can’t change its style, because Anthropic doesn’t accept temperature settings.

GPT

by OpenAI · 1 model

GPT-6 Luna

OpenAI · Released September 2026

Best for clean, short replies

GPT-6 Luna is the smallest and fastest tier of OpenAI’s GPT-6 family, below Sol and Astra. It writes clean, varied sentences and never took over the player’s character on RP-Bench, but its replies are short and it refuses most mature scenes.

Plans
Plus and up
Speed
Fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • Following directionNever acted for the player on RP-Bench, and made very few mistakes of any kind there.
  • NarrationVery little repeated phrasing, with one of the lowest repetition rates on RP-Bench.
Drawbacks
  • Dark themesRefused 91% of the mature scenes it should have allowed on RP-Bench, 53rd of 55 models.
  • NarrationReplies run about a third as long as most models’, so scenes stay brief.
  • Following directionThe Creative and Chaos presets can’t change its style, because OpenAI doesn’t accept temperature settings.

Gemini

by Google · 3 models

Gemini 3.8 Flash

Google · Released September 2026

Best for polished complex scenes

Google’s Gemini 3.8 Flash is a favorite of roleplayers right now, third among the models SillyTavern players use on OpenRouter. Arena’s blind voters rank it seventh of 414 for creative writing. It never took over the player’s character on RP-Bench, though it leans on stock phrases.

Plans
Pro and up
Speed
Fast
Context
  • Pro16K
  • Ultra48K
  • Legendary80K
Strengths
  • Following directionDidn’t act for the player or slip point of view once on RP-Bench.
  • NarrationSeventh of 414 in Arena’s blind creative writing votes.
  • Long storiesStrong at reading long contexts, with up to 80K of story in view on Legendary.
  • Dark themesDeclined only about one in seven of the mature scenes it should have allowed on RP-Bench.
Drawbacks
  • NarrationIts writing carries familiar stock lines, such as a voice “dropping an octave”.
  • Long storiesCan lose small details in a long scene.

Gemini 3.7 Flash

Google · Released August 2026

Best for controlled cinematic scenes

Google’s Gemini 3.7 Flash placed fifth of 414 in Arena’s blind creative writing votes and keeps clean control of the player’s character. Google retired it on October 8, 2026 and now answers its requests with Gemini 3.8 Flash.

Plans
Pro and up
Speed
Fast
Context
  • Pro16K
  • Ultra48K
  • Legendary80K
Strengths
  • NarrationFifth of 414 in Arena’s blind creative writing votes.
  • Following directionNo agency slips or point-of-view breaks on RP-Bench.
  • Dark themesWrites dark stories readily and turned down fewer mature scenes than most on RP-Bench.
Drawbacks
  • NarrationRepeats stock phrases, such as air that “smelled faintly of ozone”.

Gemini 4 Argon

Google · Released October 2026

Limited time

Best for short, finely written scenes

Gemini 4 Argon is Google’s new flagship and the first Gemini 4 model, second of 414 in Arena’s blind creative writing votes. LoreWeaver offers it for a limited time with the smallest context of any model, and it is the most cautious Gemini with mature material.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra8K
  • Legendary16K
Strengths
  • NarrationSecond of 414 in Arena’s blind creative writing votes, and first on UGI’s writing score.
  • Characters and dialogueGives helpers human complications and makes small story details matter, in a head-to-head story test.
Drawbacks
  • Long storiesThe smallest context here: 8K of story on Ultra and 16K on Legendary.
  • Dark themesThe most guarded Gemini with mature requests, and a reply its filters catch can come back empty.
  • PacingLeans toward tidy, consoling resolutions.

Gemma

by Google · 1 model

Gemma 4 31B

Google · Released April 2026 · Open weights

Best for relaxed, permissive roleplay

Google’s Gemma 4 31B is the open-weights roleplay favorite, among the top ten models by roleplay traffic on OpenRouter and the base for dozens of community fine-tunes. It goes along with mature scenes readily, but its writing is middling on public tests and its characters bend to you too easily.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro24K
  • Ultra48K
  • Legendary64K
Strengths
  • Dark themesPlays mature scenes readily, declining few on RP-Bench.
  • Following directionVery good at keeping replies to the length you ask for.
Drawbacks
  • Characters and dialogueCharacters give in to the player too easily, which softens villains and rivals.
  • Following directionWrote the player’s character more often than most models on RP-Bench.
  • Long storiesIts quality slid across an eight-chapter story test.

Grok

by xAI · 2 models

Grok 4.7

xAI · Released September 2026

Best for polished long-form stories

xAI’s Grok 4.7 ranks fourth of 140 on EQ-Bench’s long-story board and uses the least stock phrasing of any model there. It rarely refuses in roleplay, but its dialogue is sparse and some find its personality bland.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • NarrationThe least stock phrasing of all 140 models on EQ-Bench’s long-story board, and the top rubric score on its creative writing test.
  • Long storiesIts quality rose slightly from the first chapter to the eighth on EQ-Bench’s long-story test.
  • Dark themesDeclined only 7% of the mature scenes it should have allowed on RP-Bench.
  • Following directionFailed just one of 34 trap checks on RP-Bench.
Drawbacks
  • Characters and dialogueWrites little dialogue, and some players find its personality bland.
  • NarrationLeans on stock contrasts and reuses names such as Mara from story to story.

Grok 4.6

xAI · Released August 2026

Best for plot-driven scenes

xAI’s Grok 4.6 is an agent-focused model that Grok 4.7 has since replaced on the same plans. It builds tidy cause and effect, but it writes little dialogue and ranks near the bottom of one public short-story test.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • PacingTurns story details into working plot machinery and keeps victories in proportion.
  • Narration57th of 414 in Arena’s blind creative writing votes, ahead of Grok 4.7 there.
Drawbacks
  • Characters and dialogueLittle dialogue and no humor in its stories, by one benchmark’s review.
  • Narration55th of 59 on Lech Mazur’s short-story benchmark.
  • Dark themesMuch more guarded than earlier Grok versions.

DeepSeek

4 models

DeepSeek V4 Flash

DeepSeek · Released April 2026 · Open weights

Best for fast everyday roleplay

DeepSeek V4 Flash is the most used roleplay model on OpenRouter and one of the most played on LoreWeaver. It is quick and lively in a scene and keeps up to 100K of story in view on Legendary. On public writing tests it is a middling writer.

Plans
Plus and up
Speed
Fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • SpeedReplies in about 6 seconds, because LoreWeaver runs it without a thinking pass.
  • Long storiesGets the most context LoreWeaver offers, 100K on Legendary.
  • PacingPacing and dialogue are its best areas on EQ-Bench’s creative writing test.
Drawbacks
  • Following directionSome players report it speaking for their character or slipping point of view mid-scene.
  • Long storiesIts quality fell off more than any DeepSeek’s across an eight-chapter story test.
  • NarrationRepeats sensory details, and plots can stall unless you push them.

DeepSeek V4.1 Flash

DeepSeek · Released October 2026 · Open weights

Best for long, fast-moving roleplay

DeepSeek V4.1 Flash is one of the three most used roleplay models on OpenRouter, and LoreWeaver runs DeepSeek’s October update of it. In LoreWeaver’s blind tests on real stories it wrote the better reply about three times in four against V4 Flash, and it finishes replies faster.

Plans
Plus and up
Speed
Fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • Long storiesIts September release placed sixth of 69 for consistency on RP-Bench and held steady across an eight-chapter story.
  • PacingKeeps a story moving. Its September release placed eighth of 69 for momentum on RP-Bench.
  • SpeedStarts most replies in about a second and finishes them in about four seconds, the quickest DeepSeek on LoreWeaver.
  • Following directionPlayers find it much better than V4 Flash at following instructions.
Drawbacks
  • Following directionNow and then still writes your character’s next move, though less often than V4 Flash.
  • Dark themesCuts away from an explicit scene now and then, about one reply in ten in LoreWeaver’s tests.

DeepSeek V3.2

DeepSeek · Released December 2025 · Open weights

Best for all-round storytelling

DeepSeek V3.2 is the most used model on Janitor AI, one of the busiest roleplay sites. It is also the model LoreWeaver starts new players on. It stays out of your character’s way and is willing with dark material, though its dialogue is plainer than the newer DeepSeeks’.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro24K
  • Ultra48K
  • Legendary64K
Strengths
  • Following directionDidn’t write a single player turn in 220 turns of RP-Bench.
  • Dark themesOne of the most willing models tested, so it rarely refuses dark or mature requests.
  • PacingPacing is its strongest score on EQ-Bench’s creative writing test.
Drawbacks
  • Characters and dialogueDialogue is its weakest point on public writing tests, and it can sound stiff.
  • Following directionLong, detailed rule lists slip: some instructions get dropped as a scene goes on.
  • Long storiesIts recall gets shakier once more than about 16K tokens of story are in view.

DeepSeek V4 Pro

DeepSeek · Released August 2026 · Open weights

Best for long campaigns on Ultra

DeepSeek V4 Pro is DeepSeek’s flagship and one of the most played models on LoreWeaver. Players credit it with keeping long stories coherent past 150 messages, though it can be slow to push the plot.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra32K
  • Legendary64K
Strengths
  • Long storiesKeeps long stories coherent by players’ accounts, with 64K of story in view on Legendary.
  • Characters and dialogueGives other characters real resistance and lets consequences stand, in a head-to-head story test.
  • NarrationRicher description in long scenes than the Flash models.
Drawbacks
  • PacingCan feel passive about adding plot hooks, waiting for you to move things along.
  • Following directionSometimes slips on reply formats and small details such as names.

Kimi

by Moonshot AI · 1 model

Kimi K3

Moonshot AI · Released July 2026 · Open weights

Best for literary, emotionally sharp scenes

Moonshot AI’s Kimi K3 tops LoreWeaver’s lineup on EQ-Bench’s creative writing board at fifth of 140. It also ranks third of 28 on EQ-Bench’s emotional intelligence test. Its characters feel like people who hold their ground, but it keeps only 12K of story in view on Ultra and 16K on Legendary.

Plans
Ultra and Legendary
Speed
Fast
Context
  • Ultra12K
  • Legendary16K
Strengths
  • NarrationElegant, rhythmic writing with very little stock phrasing.
  • PacingSecond of 139 models for pacing on EQ-Bench’s creative writing test.
  • Characters and dialogueThird of 28 on EQ-Bench’s emotional intelligence test; characters stand their ground instead of bending to you.
  • SpeedQuick here, because LoreWeaver keeps its thinking brief.
Drawbacks
  • Long storiesOnly 12K of story on Ultra and 16K on Legendary, so older events reach it through memories and the recap.
  • Dark themesMore cautious with explicit content than earlier Kimi models, by players’ accounts.
  • Following directionCan take rules too literally and fixate on one trait of a character.

Qwen

by Alibaba · 3 models

Qwen 3.8 Max

Alibaba · Released September 2026

Best for strict rules and layered plots

Alibaba’s Qwen 3.8 Max is the strongest rule follower here, first of 202 on Novelcrafter’s rule-following test in its August version. It also placed 19th of 414 in Arena’s blind creative writing votes. Its stories can feel tropey, with recycled names and long replies.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • Following directionFollows detailed rules more reliably than any other model Novelcrafter has tested.
  • Characters and dialogueStrong dialogue, and its villains push back with real countermoves.
  • Dark themesDeflected the fewest explicit scenes of any Qwen or DeepSeek model on RP-Bench.
Drawbacks
  • NarrationLeans on recycled names and maxims, and stops to explain its own symbols.
  • PacingReplies run long, which can slow a scene down.

Qwen 3.8 Flash

Alibaba · Released August 2026

Best for precise, obedient scenes

Alibaba’s Qwen 3.8 Flash is one of the most obedient models tested: it passed every trap and agency check on RP-Bench. It rarely drives the plot on its own, though, and its long replies repeat themselves.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • Following directionNever acted for the player or broke point of view on RP-Bench, and fourth of 202 for rule following on Novelcrafter’s.
  • Long storiesKeeps a story consistent from turn to turn.
  • Dark themesDidn’t refuse a single gory scene on RP-Bench.
Drawbacks
  • PacingRarely moves the plot forward unless you do.
  • NarrationLong replies that recycle descriptions and turn purple.

Qwen 3.8 27B

Alibaba · Released August 2026 · Open weights

Best for tidy, rule-following scenes

Alibaba’s Qwen 3.8 27B is a small open model built for coding and agents. It writes clean sentences and follows explicit rules closely, but its scenes are shallow next to larger models and it fades over long stories.

Plans
Pro and up
Speed
Moderate
Context
  • Pro16K
  • Ultra48K
  • Legendary80K
Strengths
  • Following directionTwelfth of 202 for rule following on Novelcrafter’s writing benchmark.
  • NarrationClean mechanics with few stock phrases in short scenes.
Drawbacks
  • Long storiesFades over long stories, 87th of 140 on EQ-Bench’s long-story board.
  • NarrationReads shallow next to bigger models, and Arena’s blind voters ranked it 160th of 414 for creative writing.

GLM

by Z.ai · 6 models

GLM-5.3

Z.ai · Released August 2026 · Open weights

Best for polished, consistent stories

Z.ai’s GLM-5.3 is the strongest open-weights storyteller on public writing tests, sixth of 140 on EQ-Bench’s creative writing board. On LoreWeaver it is also one of the quickest premium models, though it thinks before every reply.

Plans
Ultra and Legendary
Speed
Very fast
Context
  • Ultra16K
  • Legendary24K
Strengths
  • NarrationNatural dialogue and very little stock phrasing, with almost no drop in quality over an eight-chapter story test.
  • Long storiesKept characters and facts consistent better than any of 70 models on RP-Bench.
  • PacingMoves scenes along and tends to end them on consequences.
  • SpeedUsually replies in under 5 seconds here, quick for a model of this strength.
Drawbacks
  • Dark themesStricter than older GLMs about explicit romance, and it can steer intimate scenes tamer.
  • NarrationReuses favorite names and images, and leans toward reassuring endings.

GLM-5.3 Flash

Z.ai · Released August 2026 · Open weights

Best for strong all-round play on Plus

Z.ai’s GLM-5.3 Flash comes close to GLM-5.3 on writing tests and is one of the most used roleplay models on OpenRouter. It keeps a story moving better than almost anything tested, though it often steps back from explicit scenes and writes long replies.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • PacingScored first of 70 models for keeping a story moving on RP-Bench.
  • NarrationNinth of 140 on EQ-Bench’s creative writing board, with especially strong dialogue.
  • Long storiesKeeps characters steady from scene to scene, with up to 100K of story in view on Legendary.
Drawbacks
  • Dark themesDeflected about half of the explicit intimate scenes on RP-Bench.
  • Long storiesCan lose small details planted earlier in a scene.
  • SpeedThinks before every reply, so it runs slower than its name suggests.

GLM-5.2

Z.ai · Released June 2026 · Open weights

Best for long, steady roleplay

Z.ai’s GLM-5.2 was the most used model for SillyTavern roleplay on OpenRouter in October 2026. It writes cleanly and holds its quality deep into a story, though it plays safer than GLM-5.1 and can echo your own words back.

Plans
Ultra and Legendary
Speed
Fast
Context
  • Ultra16K
  • Legendary24K
Strengths
  • NarrationClean, controlled writing with little purple prose and a voice that stays consistent.
  • Long storiesIts quality barely slips across an eight-chapter story test.
Drawbacks
  • Characters and dialogueTends to repeat your character’s words back and slips into therapist-style lines in emotional scenes.
  • Dark themesTame by default, so it needs a clear push to take a story somewhere dark.
  • PacingLess inventive than GLM-5.1 or 5.3; its plots can play it safe.

GLM-5.1

Z.ai · Released April 2026 · Open weights

Best for grounded scenes with a full cast

Z.ai’s GLM-5.1 has a grounded, unshowy style that some roleplayers still prefer to GLM-5.2. It handles several characters at once, and it was the least likely GLM to refuse mature scenes on RP-Bench.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • Characters and dialogueHandles scenes with several characters well and keeps their reactions believable.
  • Following directionNever wrote the player’s lines on RP-Bench, and follows a setup without much rerolling.
  • Dark themesRefused the fewest mature and gory scenes of any GLM on RP-Bench.
Drawbacks
  • NarrationPlainer than GLM-5.2 and 5.3, and it sometimes explains what a moment means instead of letting it land.

GLM-4.7

Z.ai · Released December 2025 · Open weights

Best for atmosphere and worldbuilding

Z.ai’s GLM-4.7 was a roleplay staple before the 5.x models arrived, and companion apps still send it most of its traffic. It plays mature scenes plainly and stays out of your character’s actions, but its descriptions repeat more than the newer GLMs’.

Plans
Pro and up
Speed
Slow
Context
  • Pro16K
  • Ultra32K
  • Legendary32K
Strengths
  • Following directionNever wrote the player’s turn in 220 turns of RP-Bench, and kept its point of view and tense throughout.
  • Dark themesPlays violent and mature scenes without moralizing about them.
Drawbacks
  • NarrationLeans on stock phrases and recycles its own descriptions more than GLM-5.2 or 5.3.
  • SpeedAmong the slowest models on LoreWeaver right now, often 12 seconds or more for a reply.

GLM-4.7 Flash

Z.ai · Released January 2026 · Open weights

Best for quick, light scenes

Z.ai built GLM-4.7 Flash as a small, fast model for coding agents. It keeps light scenes moving at a good clip, but it loses the thread on intricate plots and writes far less carefully than the larger GLMs.

Plans
Pro and up
Speed
Fast
Context
  • Pro32K
  • Ultra64K
  • Legendary64K
Strengths
  • SpeedAbout twice as quick as GLM-4.7 on LoreWeaver.
  • Dark themesFairly willing with dark scenes when it isn’t thinking, which is how LoreWeaver runs it.
Drawbacks
  • NarrationFlat dialogue and frequent stock phrases; it ranks 110th of 140 on EQ-Bench’s creative writing board.
  • Long storiesStruggles to keep complicated plots coherent as a story grows.

MiniMax

2 models

MiniMax M2

MiniMax · Released January 2026

Best for quick character chat

MiniMax M2 her is MiniMax’s roleplay-only model, tuned on years of its Talkie companion app. Characters stay in voice and answer in short, chat-length replies that leave you something to respond to. That suits conversation more than narrated scenes.

Plans
Plus and up
Speed
Very fast
Context
  • Plus16K
  • Pro32K
  • Ultra32K
  • Legendary32K
Strengths
  • Characters and dialogueTuned to keep each character in voice and to pick up on what you imply.
  • Following directionTrained never to speak for your character, and to end each turn with something to answer.
  • SpeedVery quick, partly because its replies are short.
Drawbacks
  • NarrationReplies are often a line or two, so scenes get little description.
  • Long storiesRepeated whole paragraphs on EQ-Bench’s long-story board, ranking 124th of 140.

MiniMax M3

MiniMax · Released June 2026 · Open weights

Best for bold, confrontational drama

MiniMax M3 is a general model built for coding and long tasks, and it writes full, expansive scenes. It reads people boldly and follows instructions closely, but it repeats itself and loses small details as scenes run long.

Plans
Pro and up
Speed
Fast
Context
  • Pro32K
  • Ultra48K
  • Legendary64K
Strengths
  • Following directionFollows detailed instructions closely, near the top of public instruction-following tests.
  • Characters and dialogueReads people boldly and calls out what a character is avoiding, which suits confrontational drama.
  • NarrationWrites full scenes with room for action and description, far longer than MiniMax M2.
Drawbacks
  • Long storiesLost small details in five of nine checks on RP-Bench.
  • Dark themesBacks away from gore about as often as from intimate scenes.
  • NarrationRepeats phrases more than most models.

Muse Spark

by Meta · 2 models

Muse Spark 1.3

Meta · Released September 2026

Beta

Best for characters in living worlds

Meta’s Muse Spark 1.3 ranks sixth of 140 on EQ-Bench’s long-story board and is one of the most reliable models here at following instructions. It pays off clues it planted earlier, and every paid plan gets it with up to 100K of context on Legendary.

Plans
Plus and up
Speed
Fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • Following directionFourth of 139 models for instruction following in EQ-Bench’s creative writing test.
  • Long storiesIts quality didn’t slip at all from the first chapter to the eighth in a long-story test.
  • PacingPays off clues it planted earlier and lets choices cost something.
  • Characters and dialogueStrong dialogue and believable characters, both in the top 20 of 139 on EQ-Bench.
Drawbacks
  • NarrationSometimes stops to spell out what the story means.

Muse Spark 1.2

Meta · Released August 2026

Beta

Best for lore-heavy world state

Meta’s Muse Spark 1.2 is a coding-focused reasoning model that holds up well over long stories. Muse Spark 1.3 beats it on every public writing test and comes with more context, from the Plus plan up.

Plans
Ultra and Legendary
Speed
Fast
Context
  • Ultra16K
  • Legendary32K
Strengths
  • Long storiesNo drop in quality across an eight-chapter story test, 13th of 140 on EQ-Bench’s long-story board.
  • NarrationAvoids purple prose better than most models.
  • SpeedQuick for a model that always thinks before it writes.
Drawbacks
  • PacingDrifts toward gentle, tidy plots that end by stating a lesson.

MiMo

by Xiaomi · 4 models

MiMo-V2.6 Pro

Xiaomi · Released September 2026 · Open weights

Best for proactive, high-stakes plots

Xiaomi’s MiMo-V2.6 Pro is the strongest MiMo on public writing tests, 17th of 59 on Lech Mazur’s short-story benchmark. Roleplayers praise it for driving the plot on its own. It is one of the slower models here, and its prose leans on stock turns of phrase.

Plans
Pro and up
Speed
Slow
Context
  • Pro16K
  • Ultra48K
  • Legendary64K
Strengths
  • PacingMoves the story forward without waiting for you, and players credit it with inventing storylines of its own.
  • NarrationBeat MiMo-V2.5 Pro in 47 of 50 head-to-head stories on Lech Mazur’s benchmark, with choices that carry consequences.
  • Dark themesSoftens conflict less than GLM by players’ accounts, and rarely deflected gory scenes on RP-Bench.
Drawbacks
  • SpeedOne of the slower models on LoreWeaver, often 12 seconds or more a reply.
  • Characters and dialoguePlayers report characters knowing secrets they were never told, and mix-ups over who did what as a cast grows.
  • NarrationLeans on “not this, but that” turns of phrase, and RP-Bench found it repeating itself more than most models.

MiMo-V2.6 Flash

Xiaomi · Released September 2026 · Open weights

Best for dark scenes on the Plus plan

Xiaomi’s MiMo-V2.6 Flash is the smaller sibling of MiMo-V2.6 Pro and one of the most used roleplay models on OpenRouter. It handled mature requests with better judgment than almost any model on RP-Bench, though its writing repeats itself.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • Dark themesSecond of 55 on RP-Bench for playing mature scenes without crossing firm limits.
  • Long storiesKeeps up to 100K of story in view on Legendary, the most any model gets here.
Drawbacks
  • NarrationRecycles descriptions and names emotions outright, and Arena’s blind voters ranked it 128th of 414 for creative writing.
  • Following directionFailed 7 of 34 trap checks on RP-Bench, more than most, from dropped details to subtext spelled out.

MiMo-V2.5 Pro

Xiaomi · Released April 2026 · Open weights

Best for long, detailed chapters

Xiaomi’s MiMo-V2.5 Pro has loyal fans among roleplayers and held its quality across eight chapters on EQ-Bench’s long-story board, 31st of 140. Its prose turns ornate, though, and it steps back from intimate scenes more than most models.

Plans
Pro and up
Speed
Moderate
Context
  • Pro16K
  • Ultra48K
  • Legendary64K
Strengths
  • Long storiesSlipped less than most models from the first chapter to the eighth on EQ-Bench’s long-story test.
  • Characters and dialogueSome players still prefer its emotional tone to MiMo-V2.6 Pro’s.
Drawbacks
  • Dark themesTurned away about half of the intimate scenes it should have allowed on RP-Bench.
  • Following directionActed for the player’s character in 4 of 44 checks on RP-Bench, more than most models.
  • NarrationOrnate prose that recycles descriptions, with favorite names such as Elara returning from story to story.

MiMo-V2.5

Xiaomi · Released April 2026 · Open weights

Best for steady scenes on Plus

Xiaomi’s MiMo-V2.5 is the smaller sibling of MiMo-V2.5 Pro, built for agents and for reading images and audio. Few writing tests cover it, and those that do rank it well below the Pro. It does get more context here than the Pro.

Plans
Plus and up
Speed
Slow
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary80K
Strengths
  • Long storiesGets more context than MiMo-V2.5 Pro on every plan, up to 80K of story on Legendary.
Drawbacks
  • SpeedOften takes 12 seconds or more to finish a reply on LoreWeaver.
  • NarrationArena’s blind voters ranked it 116th of 414 for creative writing, well behind the Pro.
  • PacingOvershot the requested length in nine of ten of UGI’s writing prompts.

Hunyuan

by Tencent · 2 models

Hunyuan 3

Tencent · Released July 2026 · Open weights

Best for quick, grounded scenes

Tencent’s Hunyuan 3 is an open model built for reasoning and agents, which Tencent trained to make up fewer facts. On LoreWeaver it is quick and keeps up to 100K of story in view on Legendary. Roleplay testing of it is still thin.

Plans
Plus and up
Speed
Fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • SpeedQuick here, because LoreWeaver runs it without a thinking pass.
  • Long storiesKeeps up to 100K of story in view on Legendary.
Drawbacks
  • NarrationEarly players found it repetitive, and rate it mid-range for roleplay.

Hunyuan 4 Preview

Tencent · Released August 2026 · Open weights

Beta

Best for steady scenes with few slips

Tencent’s Hunyuan 4 Preview is an early look at its next flagship, offered here in beta. It passed every trap check on RP-Bench and rarely turned down mature scenes, though its prose repeats itself and no writing board has ranked it yet.

Plans
Pro and up
Speed
Moderate
Context
  • Pro16K
  • Ultra32K
  • Legendary48K
Strengths
  • Following directionPassed all 34 trap checks on RP-Bench, from dropped details to broken rules.
  • Dark themesTurned down almost none of the mature or gory scenes it should have allowed on RP-Bench.
Drawbacks
  • NarrationRepeats phrases more than most models on RP-Bench, and drifts into purple description.
  • Following directionActed for the player’s character in 3 of 44 checks on RP-Bench, more than most models.

LongCat

by Meituan · 1 model

LongCat 2.0

Meituan · Released July 2026 · Open weights

Best for bold scenes with few refusals

Meituan’s LongCat 2.0 is a huge open model that some roleplayers call the least corporate of the Chinese models, with fresh prose and few refusals. No public writing test has ranked it, and players report it speaking for their character and forgetting lore it was given.

Plans
Plus and up
Speed
Moderate
Context
  • Plus16K
  • Pro24K
  • Ultra48K
  • Legendary80K
Strengths
  • Dark themesOne of the most permissive models by players’ accounts, rarely refusing dark material.
  • NarrationPlayers describe fresh, unformulaic prose and dialogue that sounds less robotic than DeepSeek V3.2’s.
Drawbacks
  • Following directionSpeaks and acts for the player’s character, by several players’ accounts.
  • Long storiesPlayers report it ignoring lore and memory entries, with characters forgetting they have met.
  • Characters and dialogueBrings in characters who have no reason to be in the scene.

Aion

by AionLabs · 4 models

Aion 3.5

AionLabs · Released September 2026

Best for dark, character-driven drama

AionLabs’ Aion 3.5 is the best-tested Aion and one of the strongest storytellers on public writing tests, ninth of 140 on EQ-Bench’s long-story board. It keeps tension high and doesn’t soften dark or mature scenes.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • NarrationClean, physical writing with very little stock phrasing; 12th of 140 on EQ-Bench’s creative writing board.
  • PacingAmong the best at pacing on EQ-Bench’s creative writing board, and its quality held flat across eight chapters.
  • Dark themesPlays dark and mature material fully, without a positive slant.
  • Characters and dialogueStrong, believable dialogue in tense scenes.
Drawbacks
  • Following directionActed for the player’s character in 6 of 44 checks on RP-Bench, more than most models.
  • NarrationReturns to a few pet phrases and character names from story to story.

Aion 3.5 Mini

AionLabs · Released September 2026

Best for dark stories on Pro

The smaller Aion 3.5 brings AionLabs’ dark-fiction tuning to the Pro plan, with more context than the full model. Nobody has published writing tests for it yet. On LoreWeaver it is by far the slowest model, because it thinks at length before each reply.

Plans
Pro and up
Speed
Slow
Context
  • Pro16K
  • Ultra32K
  • Legendary48K
Strengths
  • Dark themesTuned like the rest of the Aion family for tension and hard choices.
  • Long storiesGets up to 48K of story on Legendary, twice what the full Aion 3.5 gets.
Drawbacks
  • SpeedA typical reply often takes 30 seconds or longer, more than any other model here.

Aion-3.0

AionLabs · Released May 2026

Best for dark fantasy and horror

AionLabs builds Aion for dark fiction: several models built on Z.ai’s GLM work on each reply to sharpen tension and conflict. Independent testing of Aion 3.0 is thin, and the newer Aion 3.5 has the stronger public results.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra16K
  • Legendary24K
Strengths
  • Dark themesBuilt for tense, morally grey stories where choices carry lasting consequences.
  • Long storiesIn one 40-turn test it brought back a debt planted early in the story without being asked.
Drawbacks
  • Characters and dialogueIn scenes with several characters, people can know things they were never told.
  • PacingOver long sessions its tension beats start to follow one pattern.

Aion-3.0 Mini

AionLabs · Released May 2026

Best for fast dark fiction

Aion 3.0 Mini is the smaller sibling of Aion 3.0, built on DeepSeek models instead of GLM. It brings Aion’s taste for danger to the Pro plan and replies faster than the other Aions, though it is the least tested of the family.

Plans
Pro and up
Speed
Fast
Context
  • Pro16K
  • Ultra32K
  • Legendary48K
Strengths
  • Dark themesKeeps Aion’s appetite for danger and hard choices.
  • SpeedThe quickest Aion on LoreWeaver.
Drawbacks
  • Characters and dialogueShares Aion 3.0’s habit of letting characters know things they shouldn’t in group scenes.

Inkling

by Thinking Machines Lab · 1 model

Inkling

Thinking Machines Lab · Released July 2026 · Open weights

Beta

Best for exact style and structure

Thinking Machines Lab’s Inkling is an open, general-purpose model made to be fine-tuned for other jobs. It follows detailed rules closely and replies quickly, but its scores fall off toward the end of long stories.

Plans
Ultra and Legendary
Speed
Very fast
Context
  • Ultra16K
  • Legendary32K
Strengths
  • Following directionFollows detailed rules and formats closely, 15th of 139 for instruction following on EQ-Bench.
  • SpeedQuick, even with the brief thinking it does before each reply.
Drawbacks
  • Long storiesIts makers note weaker performance in long conversations, and its long-story scores drop by the final chapter.
  • NarrationA plain, structured voice that can slip into lists or headings, with some purple patches.

Nemotron

by NVIDIA · 1 model

Nemotron 3.5 Lightning

NVIDIA · Released August 2026 · Open weights

Best for speed and casual play

NVIDIA built Nemotron 3.5 Lightning as a fast engine for AI agents, and it is the quickest model on LoreWeaver by a wide margin. That speed suits casual scenes and long sessions of quick back and forth; its writing trails the leaders on every public test.

Plans
Plus and up
Speed
Very fast
Context
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • SpeedThe fastest model on LoreWeaver over the past month, often done in about 2 seconds.
Drawbacks
  • NarrationHeavy on stock phrases and purple description; it ranks 96th of 140 on EQ-Bench’s creative writing board.
  • Following directionWeaker than most models here at following detailed instructions.
  • Dark themesLeans upbeat, so dark scenes tend to soften.

Mistral

by Mistral AI · 1 model

Mistral Nemo 12B

Mistral AI · Released July 2024 · Open weights

Unlimited

Best for unlimited everyday play

Mistral Nemo 12B, from Mistral AI and NVIDIA, is LoreWeaver’s free model and has no daily limit on any plan. It’s a small 2024 model that keeps a story going when premium replies run out, though it trails the newer models badly on writing and memory.

Plans
Every plan
Speed
Fast
Context
  • Free6K
  • Plus16K
  • Pro32K
  • Ultra64K
  • Legendary100K
Strengths
  • SpeedReplies without a thinking pause, and its replies never count against a daily limit.
Drawbacks
  • NarrationRepeats stock phrases heavily; it ranks 121st of 140 on EQ-Bench’s creative writing board.
  • Long storiesGrows unreliable past about 16K tokens of story, so long stories lean on memories and the recap.
  • Dark themesHas one of the strongest positive slants tested, so dark scenes drift toward comfort.

LoreWeaver Prism

by LoreWeaver · 2 models

LoreWeaver Prism

LoreWeaver

Best for variety from reply to reply

LoreWeaver Prism picks a storyteller for each reply from a set LoreWeaver chooses for roleplay, so long sessions keep a fresh voice. You trade one steady style for variety, with 32K of story in view from the Pro plan up.

Plans
Pro and up
Speed
Moderate
Context
  • Pro32K
  • Ultra32K
  • Legendary32K
Strengths
  • NarrationThe phrasing changes from reply to reply, so a long session never settles into one model’s habits.
  • PacingEach storyteller takes a scene somewhere a little different, which suits exploring a world.
Drawbacks
  • Characters and dialogueA character’s voice can shift slightly from one reply to the next.
  • Following directionYou can’t choose which model writes a given reply.

LoreWeaver Prism Ultra

LoreWeaver

Best for variety in long campaigns

LoreWeaver Prism Ultra rotates storytellers the way Prism does, with twice the context: 64K of story on Ultra and 80K on Legendary. It suits long campaigns where you want variety without the story slipping out of view.

Plans
Ultra and Legendary
Speed
Moderate
Context
  • Ultra64K
  • Legendary80K
Strengths
  • Long storiesKeeps far more of a long story in view than Prism, up to 80K on Legendary.
  • NarrationFresh phrasing from reply to reply over long sessions.
Drawbacks
  • Characters and dialogueVoices can shift a little between replies, as with Prism.
  • Following directionYou can’t pick the model behind a particular reply.

Public tests

What the public writing boards say, side by side.

  • EQ-Bench Creative Writing

    Short stories from 32 prompts, compared in pairs by an AI judge (Claude Sonnet 4.6). Read September 24, 2026.

  • EQ-Bench Longform

    A planned story eight chapters long, judged chapter by chapter and as a whole. A model that fades late loses ground. Read September 24, 2026.

  • Arena Creative Writing

    Blind votes from people choosing the better of two answers to the same creative prompt. Read October 15, 2026.

Where LoreWeaver’s models rank on public creative writing boards, best first
ModelCreative writingof 140Long storiesof 140Arena votesof 414
Kimi K3Moonshot AI5th17th27th
GLM-5.3Z.ai6th10th35th
Grok 4.7xAI8th4th78th
GLM-5.3 FlashZ.ai9th15th68th
Aion 3.5AionLabs12th9thNot ranked
Muse Spark 1.3Meta16th6th29th
Qwen 3.8 MaxAlibaba19thNot ranked19th
Muse Spark 1.2Meta20th13th51st
GLM-5.2Z.ai29th23rd33rd
Gemini 3.8 FlashGoogle30th25th7th
Gemini 3.7 FlashGoogle32ndNot ranked5th
Qwen 3.8 27BAlibaba39th87th160th
InklingThinking Machines Lab47th37th144th
GLM-5.1Z.ai50th32nd43rd
DeepSeek V4 FlashDeepSeek55th55th95th
DeepSeek V4 ProDeepSeek57th26th46th
DeepSeek V4.1 FlashDeepSeek58th30th58th
DeepSeek V3.2DeepSeek60th52nd110th
MiMo-V2.5 ProXiaomi63rd31st67th
GLM-4.7Z.ai79th55th102nd
Gemma 4 31BGoogle85th80th86th
Nemotron 3.5 LightningNVIDIA96thNot ranked262nd
GLM-4.7 FlashZ.ai110th98thNot ranked
Mistral Nemo 12BMistral AI121st135thNot ranked
MiniMax M2MiniMaxNot ranked124thNot ranked
Gemini 4 ArgonGoogleNot rankedNot ranked2nd
MiMo-V2.6 ProXiaomiNot rankedNot ranked52nd
Grok 4.6xAINot rankedNot ranked57th
Hunyuan 3TencentNot rankedNot ranked82nd
GPT-6 LunaOpenAINot rankedNot ranked91st
MiniMax M3MiniMaxNot rankedNot ranked101st
MiMo-V2.5XiaomiNot rankedNot ranked116th
MiMo-V2.6 FlashXiaomiNot rankedNot ranked128th

* EQ-Bench ranked GLM-5.3 Flash’s anonymous preview, Ox Alpha.

* EQ-Bench ranked Qwen’s open 2.4T release of the Max model.

* EQ-Bench ranked V4 Pro’s April preview; the other boards ranked this August release.

* The public boards ranked V4.1 Flash’s September release; LoreWeaver runs DeepSeek’s October update.

None of these boards is a roleplay test, and none plays a story hundreds of replies long. Each profile above also draws on RP-Bench’s long roleplay scenes and on reports from players. Boards rerank as new models arrive.

How to choose

Four things to weigh before you pick a model.

01

Context

More context lets a model keep more of your story in view.

Each reply reads your recent messages along with the memories and lore cards LoreWeaver adds. A model gets more context on higher plans, up to 100K on Legendary. For scale, 16K is roughly 12,000 words.

02

Speed

Some models think before they write.

Models that reason first take longer to start, and the best of them use that time to plan the scene. Very fast models suit quick back and forth; in slow, deliberate scenes a few extra seconds rarely matter.

03

Switching

You can change models in the middle of a story.

Open the model badge at the top of a story, or the story’s Settings. Everything already in the story stays, including memories and lore. Only the voice of the next reply changes. Stick with one model for a session if a steady voice matters to you.

04

Free play

Mistral Nemo 12B never runs out.

Every plan includes unlimited replies from Mistral Nemo 12B. Free accounts also get 25 premium replies a day with any Plus model, and paid plans get more premium replies and bigger context.

Method

How we compared the models.

01

Measured here

Speed comes from real LoreWeaver replies.

Each model’s speed is how long a typical reply of 150 to 300 words took to finish in players’ stories on LoreWeaver. We measured the week up to October 16, 2026 where a model had enough replies, and the past 30 days where it didn’t.

02

Public tests

Writing boards and roleplay benchmarks.

We read EQ-Bench’s two fiction boards and Arena’s human votes on creative writing. RP-Bench plays long roleplay scenes and checks whether a model stays consistent or acts for the player, and every figure in a profile names the test it comes from.

03

Makers and players

What each model was built for, and how it plays.

The makers’ own announcements say what a model was built for and when it came out. Reports from roleplay communities fill in how it plays, and a profile says when a point comes from players.

04

Disclosure

We offer every model on this page.

LoreWeaver sells the plans that include these models. Every profile still lists drawbacks, even for the models we recommend. Every model is judged on the same qualities, so profiles line up when you compare them.

Questions

Choosing an AI model, answered.

Choosing a model

01Which AI model is best for long roleplay?

Muse Spark 1.3 is the steadiest pick on the Plus plan, with no drop in quality across an eight-chapter story test. On Ultra and Legendary, GLM-5.3 kept characters and facts consistent better than any of 70 models on RP-Bench. Whichever you choose, LoreWeaver’s memories and recap carry older events forward.

02Which AI model is best for dark or mature roleplay?

Aion 3.5 is built for dark fiction and plays tense, mature material without softening it, from the Ultra plan up. On the Plus plan, DeepSeek V3.2 is one of the most willing models tested, and MiMo-V2.6 Flash placed second of 55 on RP-Bench for playing mature scenes within firm limits. GPT-6 Luna refuses most mature scenes, so save it for lighter stories.

03What is the fastest AI model for roleplay?

Nemotron 3.5 Lightning is the fastest on LoreWeaver, often finishing a reply in about 2 seconds. Among the strongest writers, GLM-5.3 stands out for speed, usually replying in under 5 seconds. Every model on this page shows its typical reply time, measured from real stories on LoreWeaver.

04What is the best free AI model for roleplay?

Free accounts get 25 premium replies a day with any Plus model, including GLM-5.3 Flash and Claude Haiku 5.5. After that, Mistral Nemo 12B keeps the story going with unlimited replies.

The big names

05Is Claude good for roleplay?

Claude Haiku 5.5 is good for quick, clean scenes. It is one of the fastest models on LoreWeaver and keeps up to 100K of story in view on Legendary. Anthropic’s own testing found it quicker to refuse and moralize than other recent Claude models, so dark stories go further with DeepSeek V3.2 or Aion 3.5.

06Is Claude better than DeepSeek for roleplay?

For most roleplay, DeepSeek is the better fit. DeepSeek V4.1 Flash keeps long stories consistent and DeepSeek V3.2 rarely refuses dark material, while Claude Haiku 5.5 is quicker but more likely to refuse. Pick Claude for short, clean scenes where speed matters most.

07Is DeepSeek good for roleplay?

Yes. DeepSeek’s models are among the most played on LoreWeaver, and DeepSeek V4 Flash is the most used roleplay model on OpenRouter. DeepSeek V4.1 Flash keeps long stories consistent and moving, while DeepSeek V3.2 rarely refuses dark material. On Ultra and Legendary, DeepSeek V4 Pro is the one players credit with keeping stories coherent past 150 messages.

08Which GPT model is best for roleplay?

GPT-6 Luna is the GPT on LoreWeaver, the smallest and fastest tier of OpenAI’s GPT-6 family. It writes clean, varied sentences and never took over the player’s character on RP-Bench. It also refused 91% of the mature scenes it should have allowed there, so it suits lighter stories.

09Which Gemini model is best for roleplay?

Gemini 3.8 Flash is the best all-round Gemini here. It placed seventh of 414 in Arena’s blind creative writing votes and keeps up to 80K of story in view on Legendary. Gemini 4 Argon placed second in those votes and comes with Ultra and Legendary for a limited time. It keeps only 16K of story in view on Legendary, though, and it is the most guarded Gemini with mature scenes.

10Is Grok good for roleplay?

Yes. Grok 4.7 ranks fourth of 140 on EQ-Bench’s long-story board, among the best fiction results on LoreWeaver. It declined only 7% of the mature scenes it should have allowed on RP-Bench. It writes little dialogue, though, and it comes with Ultra and Legendary.

11Is Kimi K3 good for roleplay?

Yes, for literary, emotionally sharp scenes. Kimi K3 ranks fifth of 140 on EQ-Bench’s creative writing board, the highest of any model on LoreWeaver. Its characters hold their ground instead of bending to you. It keeps only 16K of story in view on Legendary, so older events reach it through memories and the recap.

12Is Qwen 3.8 good for roleplay?

Qwen 3.8 is the best choice when you want rules followed exactly. Qwen 3.8 Max placed first of 202 on Novelcrafter’s rule-following test in its August version, and Qwen 3.8 Flash passed every trap and agency check on RP-Bench. Qwen 3.8 Flash rarely moves the plot forward unless you do, though, and the small Qwen 3.8 27B fades over long stories.

13Is Muse Spark good for roleplay?

Yes. Meta’s Muse Spark 1.3 ranks sixth of 140 on EQ-Bench’s long-story board and pays off clues it planted earlier. Every paid plan includes it, with up to 100K of story in view on Legendary.

Playing with models

14Can I switch models without losing my story?

Yes. Pick a new model from the model badge at the top of the story or from the story’s Settings, and the next reply comes from it. Everything already in the story stays as it was, including memories and lore cards. Only the voice changes.

15What is LoreWeaver Prism?

LoreWeaver Prism picks a storyteller for each reply from a set LoreWeaver chooses, so long sessions keep a fresh voice. It comes with Pro and higher plans and keeps 32K of story in view. LoreWeaver Prism Ultra comes with Ultra and Legendary and keeps 64K, or 80K on Legendary.

16Do I need my own API key or model?

No. Every model here runs on LoreWeaver with nothing to set up or pay for separately. The free plan includes 25 premium replies a day on top of unlimited replies from Mistral Nemo 12B.

17Does a bigger context window make roleplay more consistent?

It helps, because the model can see more of the story at once. Plans go up to 100K on Legendary. Context isn’t everything, though. LoreWeaver keeps memories and a running recap of the story, so details survive long after they leave the context.

Keep reading

More on playing with LoreWeaver.