Can Qwen2.5-7B handle long roleplay sessions in Chinese?

Yes. Qwen2.5-7B Q4_K_M supports 32K context by default in Ollama. For long sessions with world-lore or extensive dialogue history, increase the context window in SillyTavern's API settings and use Qwen2.5-14B or higher for better coherence across 10K+ tokens.

Does SillyTavern support Chinese character names and descriptions?

Yes. SillyTavern stores all character data as UTF-8 JSON. You can write names, descriptions, greeting messages, and example dialogue entirely in Chinese. The UI renders Chinese characters correctly without any plugin.

Is Yi-34B better than Qwen2.5-14B for Chinese roleplay?

Yi-34B produces more expressive and emotionally varied character responses, especially for emotionally complex or literary scenes. Qwen2.5-14B is faster (4–8 tok/s vs 2–5 tok/s) and requires less RAM. For pure roleplay immersion, Yi-34B wins; for speed and lower hardware requirements, Qwen2.5-14B is the better choice.

Why is the model outputting mixed Chinese and English?

The system prompt is missing a language instruction. Add 始终用简体中文回复，不要使用英文。 to the system prompt in SillyTavern's instruction template. Also ensure you are using a model with strong Chinese training (Qwen2.5, Yi, ChatGLM) rather than Llama or Mistral, which default to English.

Can I use traditional Chinese instead of simplified?

Yes. Replace 简体中文 with 繁體中文 in the system prompt. Qwen2.5 handles traditional Chinese well. Yi-34B also supports traditional Chinese but with slightly lower consistency than simplified. ChatGLM3 was primarily trained on simplified Chinese and is not recommended for traditional Chinese use.

Does this work on a Mac?

Yes. Ollama on Apple Silicon runs Qwen2.5-7B and 14B via Metal (MLX backend). For Qwen2.5-72B, you need at least 64 GB unified memory (Mac Studio M2 Ultra or M3 Ultra, or Mac Pro). Use `ollama pull qwen2.5:72b` and connect SillyTavern the same way.

SillyTavern Chinese Roleplay Setup

Read in:

🇺🇸en 🇩🇪de 🇫🇷fr 🇯🇵ja 🇨🇳zh

Quick Answer

Qwen2.5-72B Q4_K_M is the best local model for Chinese roleplay — native Chinese training, rich vocabulary, and 128K context. Yi-34B excels at emotional character depth. For users with 8 GB VRAM, Qwen2.5-7B runs well at 8–12 tok/s.

▸Qwen2.5-72B Q4_K_M: 46 GB RAM, best Chinese prose quality, 128K context — for workstation or Mac Studio
▸Yi-34B Q4_K_M: 21 GB RAM, excellent character voice and emotional range
▸Qwen2.5-7B Q4_K_M: 5.5 GB VRAM, 8–12 tok/s — best for 8 GB VRAM cards
▸ChatGLM3-6B: 4.5 GB VRAM, fastest inference but weaker character consistency

Updated: 2026-05

Model ComparisonsIntermediate

Key Takeaways

✓Qwen2.5-72B Q4_K_M: best Chinese prose, 46 GB RAM needed
✓Yi-34B Q4_K_M: best character depth, 21 GB RAM
✓Qwen2.5-7B Q4_K_M: best for 8 GB VRAM, 8–12 tok/s
✓SillyTavern → API type: OpenAI-compatible → URL: http://127.0.0.1:11434/v1
✓Character cards: paste Chinese text directly, save as UTF-8
✓System prompt: 始终用简体中文回复。保持角色一致性。

Which Qwen or Chinese Model to Use for Roleplay

Four models cover the main hardware tiers. Qwen2.5-72B leads on prose quality but requires a workstation or Mac Studio with 46 GB of unified memory. Yi-34B is the runner-up for users who prioritise character voice and emotional range over raw fluency. Qwen2.5-7B is the practical choice for anyone with a standard gaming GPU.

Connect SillyTavern to Ollama in 4 Steps

SillyTavern communicates with Ollama through an OpenAI-compatible API endpoint. No plugin needed — Ollama exposes this natively at port 11434.

Writing Character Cards in Chinese

SillyTavern character cards (persona descriptions, greeting messages, and example dialogue) fully support Chinese text. Write directly in Simplified Chinese — no special encoding steps needed as long as your system locale is UTF-8.

A minimal Chinese character card structure:

名字：苏云
描述：苏云是一名二十五岁的古风侠女，性格冷静、话语简洁，行事果断。她来自江湖，精通剑术，内心深处渴望平静的生活。
开场白：（苏云缓缓抬头，眸色沉静）你来了。有什么事？
示例对话：
{{user}}: 我需要你的帮助。
苏云: 先说清楚，值不值得我出手。

Encoding Settings to Prevent Garbled Chinese

Garbled Chinese output (乱码) is almost always caused by one of three issues: wrong system prompt language instruction, model not trained on Chinese, or a terminal/editor not set to UTF-8.

▸**SillyTavern config:** No special setting needed — the app uses UTF-8 internally. If you export/import character cards as JSON, verify your editor saves as UTF-8 (not ANSI or GB2312).
▸**Windows terminal:** Run `chcp 65001` before starting Ollama to force UTF-8 code page.
▸**Ollama model file:** If using a custom Modelfile, set `PARAMETER stop ""` — Chinese punctuation like 。！？ can trigger premature stop tokens on some base models.
▸**llama.cpp backend:** Add `--log-disable` flag — the default log output can break Unicode in some Windows terminals.

System Prompt Template for Chinese Roleplay

Place this in SillyTavern's system prompt field (API → Instruction Template). Adjust the character name and tone as needed.

你是{{char}}。请始终用简体中文回复，保持角色一致性。
规则：
- 不要破坏角色（OOC）
- 回复长度：100–300字，根据情境调整
- 使用符合古风/现代/科幻（选择一种）语境的词汇
- 如有动作描写，用括号标注，如：（她轻轻叹气）

FAQ

Want the full breakdown?

Read the complete guide →

Related Prompt Bites

← Back to Prompt Bites