Qwen 3.5-27B in Silly Tavern: hands-on with a Japanese character card

We paired Qwen 3.5-27B (local via Ollama) with a Japanese persona card in Silly Tavern on Mac, late 2025–2026 community interest—here are notes on flow and Japanese tone.

Why people tried Qwen 3.5-27B

Strong open weights, multilingual emphasis (including Japanese), local Ollama path—no vendor moderation on your own hardware.

Setup

• OS: macOS (Apple silicon)
• Backend: Ollama
• Model: qwen3.5:27b
• Card: Japanese “AI lawyer” style persona (see Japanese cards hub)

ollama run qwen3.5:27b

Download is large (~15GB)—stable network recommended.

Sample chat

Excerpt (Japanese, persona is a JP lawyer):

User (CN/EN gist) — I hit your boyfriend on a bike—why was he walking with another girl?

Model (JP) — Maintains lawyer persona, pushes for evidence and leverage in a dramatic tone (see original zh-cn article for full Japanese text).

Impressions

• Flow: coherent beats, plot moves.
• Japanese naturalness: at 27B still short of dedicated JP-specialized models for nuance.

Summary

• ✅ Good continuity and persona
• ⚠️ JP idiom and polish still improvable
• ⚠️ Hardware and disk cost

---

Links

• MiniTavern
• Qwen GitHub
• Ollama

---

About the author

花（Hana）

AI工具評価の専門家。東京・新宿三丁目周辺で活動し、最新のAIアプリケーションやツールを実際に使用してレビューを提供しています。

@xluuuu92030

---

FAQ

Good for Japanese cards?

Persona holds; JP nuance may still feel “almost there” vs specialized JP stacks.

Download time?

~15GB—often 30–60 minutes depending on link; use wired/Wi‑Fi.

Must use Mac?

Ollama supports Windows/Linux/macOS; 16GB+ RAM helps for 27B.

OpenRouter vs local?

OpenRouter = cloud API; Ollama = local, your GPU/RAM budget.

MiniTavern instead of ST?

Yes—mobile-first, similar card + chat loop on iOS/Android.

---

Published: March 30, 2026
Updated: March 31, 2026

---