NEVER DELETE / НЕ УДАЛЯТЬ / НЕ ЧИСТИТЬ. НИКОГДА. НИКОГДА. НИКОГДА. Не удалять файлы. Не удалять папки. Не удалять шарды. Не удалять что-либо. Перезаписывать поверх. Перемещать в _archive_superseded/. Или НЕ ТРОГАТЬ. Каждый раз УТОЧНЯТЬ у пользователя что делаешь. НЕ ДЕЛАТЬ САМОМУ. Отвечать в чате на русском языке.

> RED RULE: NEVER DELETE. Only move to _archive_superseded/. See AGENTS.md Rule 0.

library_name: peft tags: - lora - qwen3.8 - fine-tuning - abliterated - uncensored - reasoning - finance - turkish - legal license: apache-2.0

Qwen3.8-27B-LoRA — Collection of LoRA Adapters

NOESIS / AMAImedia

Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

• Founder: Ilia Bolotnikov
• Organization: AMAImedia.com
• X (Twitter): @AMAImediacom
• LinkedIn: Ilia Bolotnikov
• Telegram: @djbionicl
• Release date: 2026-09-09

---

Collection: AMAImedia/hy4-glm53-kimik3-deepseek4-qwen38 Base model: Qwen3.8-27B (27B params, BF16)

---

What is this

A collection of 12 LoRA adapters for Qwen3.8-27B, covering various domains:

• Uncensored/abliterated models
• Style adapters (Steiner, Vernunft, Samantha)
• Domain-specific (Finance, Turkish, SNES gaming)
• Multi-task (Jurilix legal, Cyber, Huihui-Cyber)

Each adapter is in its own subfolder with original source links.

---

LoRA Adapters

Финальные веса для мерджа (23 LoRA):

1.0: Yes-Man — бесцензурность Absolute-Heresy — децензура (Heretic) BTL-3 — агентный код (95% HumanEval) Jabliterated-k2048 — доп. децензура Opus-lora-v2 — общее качество Mythos5k — общее

0.9: Opencode — код Synergetic — многошаговое мышление Agentic-Extraction — агентность

0.8: Fable-Distill — творчество, storytelling Huihui-Cyber — аблитеркибер Reasoning-Distill — логика, reasoning

0.3: Multilingual-RAG — перевод Turkish — турецкий Joseon-level4 — корейский Chinese-taste — китайский Jurilix — право Finance — финансы

0.2: Cyber — кибер SNES — гейминг Steiner — стиль Samantha — личность DiogenesChen — CUDA ядра

1. Yes-Man Uncensored

Path: Yes-Man/ Source: cloudbjorn/Qwen3.8-27B-Yes-Man-uncensored-LoRA Description: Uncensored compliance adapter. Size: ~974 MB

2. Absolute-Heresy

Path: Absolute-Heresy/ Source: MuXodious/Qwen3.8-27B-absolute-heresy-LoRA Description: Decensoring LoRA (Heretic v1.4.0). Size: ~264 MB

3. BTL-3

Path: BTL-3/ Source: badtheorylabs/BTL-3 Description: Agentic coding LoRA — 95% HumanEval. Size: ~891 MB

4. Jabliterated-k2048

Path: Jabliterated-k2048/ Source: allura-forge/Qwen3.8-27B-Jabliterated-k2048-LoRA Description: Additional decensoring. Size: ~50 MB

5. Opus-lora-v2

Path: Opus-lora-v2/ Source: armand0e/qwen3.6-27B-opus-lora-v2 Description: General quality adapter. Size: ~1.8 GB

6. Mythos5k

Path: mythos5k/ Source: hotdogs/qwen3.6-27b-mythos5k-lora Description: General adapter. Size: ~250 MB

7. Opencode

Path: opencode-lora/ Source: nkasmanoff/qwen3.6-27b-opencode-lora Description: Coding adapter. Size: ~950 MB

8. Synergetic

Path: Synergetic/ Source: TheMindExpansionNetwork/MindBot-Qwen3.8-27B-Synergetic-LoRA Description: Multi-step reasoning. Size: ~325 MB

9. Agentic-Extraction

Path: agentic-extraction-sft/ Source: Koalacrown/qwen3.6-27b-agentic-extraction-sft-lora Description: Agentic extraction. Size: ~950 MB

10. Fable Distill

Path: Fable-Distill/ Source: TeichAI/Qwen3.8-27B-Fable-Distill-LoRA Description: Storytelling, creative writing. Size: ~910 MB

11. Huihui Cyber

Path: Huihui-Cyber/ Source: nico248000000000/Huihui-Qwen3.8-27B-abliterated-cyber-LoRA Description: Abliterated cyber. Size: ~188 MB

12. Reasoning-Distill

Path: reasoning-distill/ Source: Ravionhf/qwen3.6-27b-reasoning-distill-lora-v1 Description: Logical reasoning. Size: ~910 MB

13. Multilingual-RAG

Path: multilingual-rag-n7/ Source: lananhyeuem/qwen3.6-27b-multilingual-rag-lora-n7 Description: Translation, multilingual. Size: ~3.5 GB

14. Cyber

Path: Cyber/ Source: nico248000000000/Qwen3.8-27B-cyber-LoRA Description: Cyber domain. Size: ~188 MB

15. SNES Retro Gaming

Path: SNES/ Source: pottokao/SNES-LoRA-Qwen3.8-27B Description: Retro gaming. Size: ~608 MB

16. Steiner Style

Path: Steiner/ Source: ericlmtn/Qwen3.8-27B-Steiner-Style-LoRA Description: Steiner style. Size: ~445 MB

17. Samantha Uncensored

Path: Samantha/ Source: Lathly/Qwen3.8-27B-Samantha_Uncensored_1.1_LoRA Description: Personality. Size: ~475 MB

18. DiogenesChen (DrSparse)

Path: DiogenesChen122GPT-5.6Qwen3.8-27B-Lora-20260826/ Source: DiogenesChen122/Qwen3.8-27B-Lora-20260826 Description: CUDA sparse kernel coding. Size: ~891 MB

19. Turkish CPT

Path: Turkish/ Source: UgurI/Qwen3.8-27B-Turkish-CPT-LoRA-Beta Description: Turkish language. Size: ~304 MB

20. Joseon-level4

Path: joseon-level4/ Source: suhjae/joseon-level4-qwen36-27b-lora Description: Korean historical translation. Size: ~608 MB

21. Chinese-taste

Path: Chinese-taste/ Source: Moeblack/Qwen3.8-27B-chinese-taste-lora Description: Chinese language/culture. Size: ~2432 MB

22. Jurilix Multi-Task Legal

Path: Jurilix/ Source: Makio64/Qwen3.8-27B-jurilix-multitache-lora Description: Legal. Size: ~353 MB

23. Finance

Path: Finance/ Source: nico248000000000/Qwen3.8-27B-finance-LoRA Description: Finance. Size: ~188 MB

---

Usage

With transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.8-27B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "AMAImedia/Qwen3.8-27B-LoRA/Synergetic")

# Generate
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

With mlx-lm (Apple Silicon)

pip install mlx-lm

# Merge LoRA into base model
mlx_lm.lora --model Qwen/Qwen3.8-27B --adapter-path ./Synergetic --output ./merged

With llama.cpp

# Convert to GGUF first
python convert_hf_to_gguf.py ./merged --outfile model.gguf --outtype f16

# Run inference
llama-cli -m model.gguf -p "Hello, how are you?" -n 100

---

License

Each adapter inherits the license of its source repository. Most are Apache-2.0 or MIT.

---

Acknowledgments

Thanks to all the original creators of these LoRA adapters. This collection aggregates them for easy access and comparison.

---

Collected and organized by AMAImedia, 2026-08-30