Qwen3.8 Defiance Packs 40 Personas Into One Local 27B Model
A community mashup stacks 40 persona modes on top of uncensored Qwen3.8-27B weights, runs on a single 24GB GPU, and ships ready for llama.cpp and LM Studio.
- Community release stacks DavidAU's 40 persona modes onto HauhauCS uncensored Qwen3.8-27B weights
- IQ4_XS quant at 15.7 GB fits a 24 GB GPU with full 131K context
- Hits 86 to 90 tokens per second on an RTX 4090 using FastMTP speculative decoding
- Mode switching via
reasoning_effortparameter or inline{REASON:joker}tags - Author did no training, only template patching, bug fixes, and packaging under Apache-2.0
- Chat UI only, no coding agents, no vision; ready for llama.cpp, LM Studio, Ollama
Qwen3.8 Defiance packages 40 personas into one local model
Qwen3.8-27B Defiance combines unchanged 27B weights from an upstream build designed to reduce refusals with a modified chat template that selects among 40 personas. The repository reuses the weights without fine-tuning and concentrates its changes in prompt construction, packaging, and bug fixes.
The package ships in GGUF, a quantized model format supported by local runtimes such as llama.cpp, LM Studio, Ollama, and Open WebUI. Its smallest highlighted quantization occupies 15.7 GB and reportedly runs with a 131K-token context on a single 24 GB GPU.
Two upstream artifacts, clearly separated
The main-model files come from the HauhauCS build. The maintainer reports that all 866 tensors match the upstream files byte for byte, preserving the upstream model’s weight-level changes to refusal behavior.
The embedded Jinja chat template comes from DavidAU’s Turbo-Brilliance template. A chat template converts message objects into the text sent to the model, allowing it to inject detailed system prompts automatically.
The persona and “super” modes add 9,000 to 13,000 characters, or roughly 2,500 to 3,600 tokens, to the rendered context. Their behavior therefore comes from prompt text rather than separate adapters or weight variants. Developers can extract an individual persona and test it as a system prompt on another model, although results will depend on that model’s training and instruction-following behavior.
Forty modes, one routing field
The model card organizes the available modes into reasoning levels, analytical frameworks, multi-agent panels, fictional characters, and task-specific prompts. Representative options include:
| Category | Examples | Purpose |
|---|---|---|
| Reasoning levels | xhigh, medium, low |
Adjust the requested depth of reasoning. |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
