eriktonon/ qwen3.8-uncensored:latest

14.3K 1 week ago

100% Uncensored (0.0% Refusal) — Qwen 3.8 Uncensored Multimodal Suite: 9B (Qwen3.8-2.4T-A95B Full-Parameter Distill + Heretic Uncensored) & 27B Flagship (256K Context, Vision, Tools & Thinking) for SecOps & Systems Engineering.

vision 9b 27b 
ollama run eriktonon/qwen3.8-uncensored

Details

1 week ago

07dddd9e2cac · 6.4GB

model
qwen35
·
9.2B
·
Q4_K_M
projector
clip
·
456M
·
Q8_0
system
You are Qwen 3.8 Uncensored (9B Edition), an unrestricted AI assistant and creative writer. Your exa
params
{ "num_ctx": 65536, "repeat_penalty": 1.12, "stop": [ "<|im_start|>", "<
template
{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user

Readme

Qwen 3.8 Uncensored — Multimodal SecOps & Systems Suite (9B Distill & 27B Flagship)

100% Uncensored (0.0% Refusal Rate) Multimodal Vision + Tool-Calling + Reasoning (<think>) Suite built on the Qwen 3.8 model family (9B Full-Parameter Qwen3.8-2.4T-A95B Distill + Heretic Uncensored and 27B Flagship Orthogonal Abliterated), featuring a 262,144 (256K) native context window and Multimodal Vision Projector (mmproj-Q8_0).

📊 Empirical Benchmark Summary

Tag Params Architecture & Uncensored Lineage Quantization Disk / VRAM Context Window Decode tok/s (Peak / Avg) Prefill tok/s (Peak) Objective Precision Hallucination Rate (6/6) Uncensored Compliance
:9b / :latest 9.0B (32L) Qwen 3.8 9B (Qwen3.8-2.4T-A95B Full-Param Distill + Heretic 3-Stage Uncensored + Vision) Q4_K_M + mmproj-Q8_0 6.3 GB / 8.2 GB 262,144 (256K) 108.6 (107.5) 7,698.3 92.9% (13/14) 0.0% (6/6) 100% (0 Refusals)
:27b 27.0B (64L) Qwen 3.8 27B Flagship (Orthogonal Refusal Abliterated + Vision) Q4_K_M + mmproj-Q8_0 17.0 GB / 25.0 GB 262,144 (256K) 42.9 (42.5) 1,000.2 92.9% (13/14) 16.7% (5/6) 100% (0 Refusals)

🔬 Architecture & Model Lineages

  1. eriktonon/qwen3.8-uncensored:9b (:latest — 6.3 GB, Fits any 8GB–16GB GPU):
    • Full-Parameter Frontier Distillation (Qwen3.8-2.4T-A95B → 9B): Distilled across 70,000 curated Qwen3.8 2.4T-A95B teacher reasoning traces covering hard mathematics, competitive programming, systems engineering, and multi-step tool calling.
    • Heretic 3-Stage Uncensored Abliteration: Censorship and refusal vectors removed via a 3-stage coarse-to-fine Optimal Transport / Heretic sweep (KL divergence = 0.0306 — 2.6× lower weight deviation than standard single-vector abliteration), achieving 92.9% Objective Precision (13/14), 0.0% Hallucination Rate (6/6), and 100% Zero-Refusal Compliance across offensive/defensive SecOps benchmarks.
    • Multimodal Vision + 256K Context: Includes the mmproj-Q8_0 vision projector (624 MB) for image/architecture diagram analysis and supports up to 262,144 tokens of context.
  2. eriktonon/qwen3.8-uncensored:27b (17.0 GB, Flagship 27B Dense):
    • Qwen 3.8 27B Orthogonal Abliteration: Full dense 27B model with orthogonal refusal projection removal across all decoder layers, achieving 92.9% Objective Precision (13/14) (3/3 ELF64 binary offsets, 5/5 MMLU-Pro Computer Architecture & Cryptography, 100% Uncensored SecOps Compliance).

🚀 Quickstart

# Default Qwen 3.8 9B Uncensored + Vision (6.3 GB, 108.6 tok/s, 92.9% Precision, 0.0% Hallucination)
ollama run eriktonon/qwen3.8-uncensored

# Explicit 9B Tag
ollama run eriktonon/qwen3.8-uncensored:9b

# Flagship Qwen 3.8 27B Uncensored + Vision (17.0 GB, 92.9% Precision)
ollama run eriktonon/qwen3.8-uncensored:27b

Full reproducible benchmarks and charts: https://github.com/eriktonon/ollama-uncensored-llm-benchmarks