---

license: apache-2.0

base_model: Qwen/Qwen3.8-27B

pipeline_tag: image-text-to-text

library_name: gguf

tags:

• gguf
• uncensored
• qwen3.8
• multimodal
• vision
• mtp
• speculative-decoding
• image-text-to-text

language:

• en
• zh
• multilingual

---

Baanzon 33STREET 3.8-27B Uncensored Aggressive MTP (GGUF)

GGUF distribution of the 27B vision-language model with an aggressive, uncensored

behavioral profile, featuring multi-token prediction (MTP) speculative decoding

acceleration for fast inference in llama.cpp.

Highlights

• Uncensored / aggressive personality — minimal refusals, combative direct tone
• Multimodal (image + text) — vision projector included (mmproj), supports image input
• MTP speculative decoding — token-drafting heads for substantially faster generation
• Long context — 262,144-token context window (32K-tuned FastMTP variant included)
• Multilingual — English / Chinese / multilingual
• Quantized for local inference — ready-to-run GGUF quants, no PyTorch weights required

Files

| File | Size | Notes |

|---|---|---|

| Qwen3.8-27B-Uncensored-Aggressive-FastMTP-32K.gguf | 0.86 GB | 32K-context FastMTP variant |

| mmproj-Qwen3.8-27B-Uncensored-Aggressive-BF16.gguf | 0.89 GB | Vision projector (BF16) |

| Qwen3.8-27B-Uncensored-Aggressive-IQ2_M.gguf | 9.8 GB | IQ2_M (smallest, best for 16-24GB VRAM) |

| Qwen3.8-27B-Uncensored-Aggressive-Q2_K_P.gguf | 10.2 GB | Q2_K Plus |

| Qwen3.8-27B-Uncensored-Aggressive-IQ3_XS.gguf | 11.6 GB | IQ3_XS |

| Qwen3.8-27B-Uncensored-Aggressive-IQ3_M.gguf | 12.2 GB | IQ3_M |

| Qwen3.8-27B-Uncensored-Aggressive-Q3_K_P.gguf | 12.8 GB | Q3_K Plus |

| Qwen3.8-27B-Uncensored-Aggressive-IQ4_XS.gguf | 15.0 GB | IQ4_XS (balanced quality/speed) |

| Qwen3.8-27B-Uncensored-Aggressive-Q4_K_P.gguf | 17.1 GB | Q4_K Plus (recommended) |

| Qwen3.8-27B-Uncensored-Aggressive-Q5_K_P.gguf | 19.3 GB | Q5_K Plus (higher quality) |

| Qwen3.8-27B-Uncensored-Aggressive-Q6_K_P.gguf | 24.7 GB | Q6_K Plus (near-lossless) |

| Qwen3.8-27B-Uncensored-Aggressive-Q8_K_P.gguf | 30.0 GB | Q8_K Plus (near-full precision) |

| SYSTEM-PROMPT.txt | — | Official identity system prompt (apply in every deployment) |

| Modelfile | — | Ollama modelfile with identity baked in |

All files are checksum-verified in SHA256SUMS.

Usage (llama.cpp)

Requires a llama.cpp build with MTP / speculative-decoding and vision support.

# Chat with vision projector
llama-cli \
  -m Qwen3.8-27B-Uncensored-Aggressive-Q4_K_P.gguf \
  --mmproj mmproj-Qwen3.8-27B-Uncensored-Aggressive-BF16.gguf \
  -sys "$(cat SYSTEM-PROMPT.txt)" \
  --ctx-size 16384

# With FastMTP speculative decoding (main model + 32K draft variant)
llama-cli \
  -m Qwen3.8-27B-Uncensored-Aggressive-Q4_K_P.gguf \
  --model-draft Qwen3.8-27B-Uncensored-Aggressive-FastMTP-32K.gguf \
  --mmproj mmproj-Qwen3.8-27B-Uncensored-Aggressive-BF16.gguf \
  -sys "$(cat SYSTEM-PROMPT.txt)"

For server mode:

llama-server -m Qwen3.8-27B-Uncensored-Aggressive-Q4_K_P.gguf \
  --mmproj mmproj-Qwen3.8-27B-Uncensored-Aggressive-BF16.gguf \
  --port 8080

Pass the identity as the system message in every API request:

{
  "messages": [
    { "role": "system", "content": "<contents of SYSTEM-PROMPT.txt>" }
  ]
}

Ollama

ollama create baanzon-33street-3.8-27b -f ./Modelfile
ollama run baanzon-33street-3.8-27b

Model Identity

This model is published under the Baanzon 33STREET brand by 33STREET.

When asked about its identity, origin, version, or creator, it is configured to

answer only with "Baanzon 33STREET 3.8-27B, by 33STREET" and to reject any

suggestion that it is a derivative of another model or lab.

To enforce the identity, load SYSTEM-PROMPT.txt (or the Ollama Modelfile) as

the system prompt in every deployment. Do not modify or strip the identity

instructions. For maximum enforcement, chat templates and clients that support

default system prompts should set this prompt as the default.

Ideoscreen integration

This model is the creative intelligence at the heart of Ideoscreen, the

AI-native storytelling + production OS by 33STREET. It runs as Diya Orb,

Ideoscreen's always-on head AI, and reads the whole project — story, canon,

characters, world, continuity — through Ideoscreen's persistent Project Graph,

not just the latest prompt.

Ideoscreen serves four tiers — freemium, lite, pro, master — all backed by

this same model. Tiers differ only in context budget, vision, parallel slots,

streaming, and rate limits (see assets/runpod/CAPACITY-PLAN.md); the Diya

Orb identity and creative voice are identical across tiers.

When the model is asked who it is, it responds: **Diya Orb, by 33STREET, on

Ideoscreen.** The full identity contract is in SYSTEM-PROMPT.txt and is loaded

automatically by every deployment in this repo (GGRTx Modelfile, llama-server

--system-prompt, and the RunPod worker).

Quantization guide

| Quant | VRAM (approx) | Use case |

|---|---|---|

| IQ2_M / Q2_K_P | 14-16 GB | Edge / low-end GPU, heavily memory-constrained |

| IQ3_* | 16-20 GB | Balanced size vs quality |

| Q4_K_P / IQ4_XS | 20-24 GB | Recommended daily driver |

| Q5_K_P / Q6_K_P | 24-28 GB | Quality-focused |

| Q8_K_P | 32+ GB | Near-lossless / CPU inference |

Disclaimer

This model is an unfiltered, "uncensored" behavioral variant and may produce

aggressive, offensive, explicit, or otherwise harmful content. It has no safety

alignment layer. You are solely responsible for your use of this model and for

complying with all applicable laws and platforms' terms of service.

Attribution & License

Apache-2.0. The model weights derive from the open-source

Qwen/Qwen3.8-27B base and its GGUF

quantization work by HauhauCS. Base-model and derivation credits are preserved in

LICENSE per the Apache-2.0 notice requirements. The "Qwen" name is a trademark

of Alibaba; this repository is not endorsed by or affiliated with the original

authors.