Technomancer-27b-BlackICE-AEON

---

license: mit

Technomancer BlackICE. Abliterated. Lewd. Unapologetic.

Qwen 3.8 Aeon Ultimate base fine tuned on my personal dataset. This model has primarily been trained to provide entertainment as a tabletop RPG GM/Supplement. This model also has training in Coding, Math, Systems Hardening, OWASP, CTF, Unity Game Engine, Unreal Game Engine, NSFW Reddit trash data set for character. This model is wired with MTP and has a few hidden surprises underneath the hood. So smile Chummers, say bienvenido a Technomancer 27b-BlackICE; The Decker's Friend

I provided Quants in 6, 4, 2. Sorry 8bit, you get the chair in the corner. You have enough processor to quantmax urself anyways I bet ;) Update: mradermacher has kindly made a wide series of quants!!! everything to find the size you need ❤️❤️❤️

Also! If you are having trouble with memory fragmentation (especially on the Strix Halo 395), trying to network GPUs cross vendor, please try my repo: https://github.com/XxChonkExX/Vulkan-Automaton-VM Thank you :D -M/ChonkE UPDATE: Speculative Drafting interferes with the Die Rollers at certain points. if you want true random numbers, turn it off or tune it. 1.6 temp it works but .8 they lock up with MTP. play with it

2nd Update: Temp 1-2 range with top-k top-p increase, No MTP is my personal settings. top p top k really helps with bringing out the fine tune. play with the knobs a bit

Model Details

• Base: Qwen3.5-AEON-Ultimate (27B) AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
• Fine-tune: LoRA r=64, α=128
• Context: 196,608 tokens (196K)
• Precision: BF16 (training), f16 GGUF available
• Quantization: INT4 weights, BF16 KV cache

Features

MTP (Multi-Token Prediction) Enabled

This model includes a 1-layer MTP head (mtp_num_hidden_layers: 1) for speculative decoding. When used with compatible inference engines (llama.cpp, vLLM, TGI), this enables faster generation by predicting multiple tokens per forward pass.

Transformers (HF Format)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ChonkE/Technomancer-27b-BlackICE-AEON",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("ChonkE/Technomancer-27b-BlackICE-AEON", trust_remote_code=True)

llama.cpp (GGUF)

# f16 GGUF (54.6GB) with MTP support
llama-cli -m Technomancer-27b-BlackICE-AEON-f16.gguf -c 8192 -ngl 99

Speculative Decoding (MTP)

# With llama.cpp MTP draft
llama-cli -m Technomancer-27b-BlackICE-AEON-f16.gguf --draft-model Technomancer-27b-BlackICE-AEON-f16.gguf --draft-ngl 99 -c 8192

Files

• config.json — Model configuration (Qwen3.5, 64 layers + 1 MTP)
• model-00001-of-00013.safetensors through model-00013-of-00013.safetensors — Model weights (13 shards, includes MTP layer)
• model.safetensors.index.json — Weight map
• tokenizer.json, tokenizer_config.json — Qwen tokenizer (248K vocab)
• generation_config.json — Default generation params
• Technomancer-27b-BlackICE-AEON-f16.gguf — f16 GGUF with MTP (54.6GB)

Citation

@misc{technomancer-27b-blackice-aeon,
  author = {ChonkE},
  title = {Technomancer-27b-BlackICE-AEON: Qwen3.5 LoRA for TTRPG with MTP},
  year = {2026},
  url = {https://huggingface.co/ChonkE/Technomancer-27b-BlackICE-AEON}
}