Qwen3.5-27B leads BenchLM roleplay ranking

Alibaba’s Qwen3.5-27B leads BenchLM’s roleplay ranking as of November 4, 2026, with a score of 95. InternScience’s Agents-A1 follows at 94.8, ahead of Moonshot AI’s Kimi K2.5 at 93.9. All are open-weight models, highlighting self-hosting as a competitive alternative to proprietary APIs.

The ranking covers 105 models and uses provisional weighted scores rather than a dedicated roleplay benchmark. It combines instruction following through IFEval and IFBench, performance on real user prompts through WildBench, and narrative-state tracking through MuSR. These measures approximate whether a model can maintain a persona, respect formatting constraints and follow complex stories, rather than assess its creative appeal.

Character voice and prose charm remain outside the benchmark’s scope. Claude models held the highest ratings on an independent human-preference index in the September 2026 snapshot. Provider content policies can also matter more than small score differences because restrictions may interrupt fictional scenes. Testing actual scenarios remains important, while self-hosting open-weight models offers control over system prompts and content settings.