BenchLM recommendation
Best LLMs for Roleplay in 2026
As of July 13, 2026, the top model in best llms for roleplay on the BenchLM leaderboard is Qwen3.5-27B with a score of 95.
Last verified: July 13, 2026
There is no dedicated roleplay benchmark, so this reporting family ranks the measurable ingredients: instruction following (IFEval, IFBench — whether the model stays in the persona and format you set), WildBench (performance on real, messy user prompts, including creative ones), and MuSR (tracking multi-step narrative state). Human preference for prose style is worth reading alongside: Claude models hold the top Arena Elo scores BenchLM tracks.
This page ranks models by a sourced proxy blend — instruction following, real-prompt quality, and narrative reasoning — because no standalone roleplay benchmark exists.
Bottom line: persona consistency is instruction following in costume — the IFEval/IFBench leaders below hold character best. For prose feel, cross-check Arena Elo and test your own scenarios.
According to BenchLM.ai, Qwen3.5-27B leads this ranking with a score of 95, followed by Agents-A1 (94.8) and Kimi K2.5 (93.9). The top three are separated by just a few points — any of them would perform well for this use case.
The best open-weight option is Qwen3.5-27B (ranked #1 with a score of 95). Open-weight models are highly competitive in this category — self-hosting is a viable alternative to proprietary APIs.
This ranking is based on provisional overall weighted scores across BenchLM.ai's scoring formula tracked by BenchLM.ai. For detailed model profiles, click any model name below. To compare two specific models head-to-head, use the "vs #" links.
How to choose
Long persona consistency?
Weight IFEval and IFBench — staying in character is instruction adherence
Natural conversational prose?
Claude models hold the top Arena Elo human-preference scores
Complex multi-character stories?
MuSR tests multi-step narrative reasoning directly
Private, uncensored-model hosting?
Open-weight rows can be self-hosted — see the local LLM guide
Full Rankings (31 models)
These rankings update weekly
Get notified when models move. One email a week with what changed and why.
Free. No spam. Unsubscribe anytime.
Key Takeaways
The top model on this sourced reporting-family slice is Qwen3.5-27B by Alibaba with an average of 95.
The best open-weight model is Qwen3.5-27B at position #1.
31 models are listed with sourced benchmark coverage in this reporting family.
Score in Context
What these scores mean
A proxy blend: instruction following measures whether the model keeps your persona, format, and constraints; WildBench samples real user prompts; MuSR tests narrative-state tracking. Together they approximate roleplay reliability.
Known limitations
No benchmark measures prose charm or character voice — the qualities roleplay users often care about most. Content policies also differ sharply between providers and matter more than a few benchmark points for this use case. Test your actual scenarios.
Best LLMs for Roleplay FAQ
What is the best LLM for roleplay?
By measurable proxies — instruction adherence, real-prompt quality, narrative tracking — the top rows of this table lead. For prose feel, Claude models hold the highest human-preference (Arena Elo) scores BenchLM tracks. The honest answer is to shortlist from this table and run your own scenarios; voice preference is personal.
What is the best local LLM for roleplay?
The strongest open-weight rows in this family can run locally — self-hosting also gives you full control over system prompts and content settings, which matters for this use case. See the local LLM rankings for what fits your VRAM tier, and the Ollama guide for pull commands.
Why is there no roleplay benchmark?
Because the target is subjective: persona charm and voice resist automated scoring. What can be measured — does the model follow persona instructions, handle real messy prompts, and track narrative state — is what this family blends. Treat it as a reliability floor, not a style ranking.
Do content policies matter more than benchmarks for roleplay?
Often yes. Providers differ sharply in what fiction they permit and how aggressively safety layers interrupt scenes, and that difference dwarfs a few benchmark points in practice. Check each provider's current usage policies for your use case; open-weight models under your own hosting sidestep the issue entirely.
Explore More
The AI models change fast. We track them for you.
A weekly brief for engineers and researchers covering new models, ranking shifts, and pricing changes.
Free. No spam. Unsubscribe anytime.