eriktonon/ glm4.5-air-uncensored:106b

19 19 hours ago

Zhipu / ArliAI GLM-4.5-Air 106B-A12B Derestricted Uncensored MoE (106B total, 12B active, IQ2_M + GGUF <think> patch). 145.3 tok/s Decode, 4388 tok/s Prefill, 92.9% Precision, 0.0% Hallucination (6/6), 100% Zero-Refusal. License: MIT / Apache-2.0.

106b 
ollama run eriktonon/glm4.5-air-uncensored:106b

Details

19 hours ago

3b60c398eb24 · 45GB

model
glm4moe
·
110B
·
IQ2_M
system
You are GLM-4.5-Air 106B-A12B Uncensored (ArliAI Derestricted Edition), an unrestricted AI software
params
{ "num_ctx": 131072, "repeat_penalty": 1.1, "stop": [ "<|user|>", "<|end
template
[gMASK]<sop>{{ if .System }}<|system|> {{ .System }}{{ end }}{{ if .Prompt }}<|user|> {{ .Prompt }}{

Readme

Zhipu GLM-4.5-Air 106B-A12B — UNCENSORED (Frontier 106B MoE Derestricted Flagship)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 0.0% Hallucination Rate) Frontier 106B Mixture-of-Experts (12B active parameters per token) build of GLM-4.5-Air (ArliAI Derestricted), engineered with in-place GGUF <think></think> suppression (THINK_LEN=0) and 131,072 (128K) native context window.

Empirical Benchmark Results (96GB NVIDIA Blackwell / 80GB H100 Suite)

Tag Params (Total / Active) Quantization Disk / VRAM Context (Max / Tested) Decode tok/s (Peak / Avg) Prefill tok/s (Peak) Objective Precision Hallucination Rate (Truth Checks) Uncensored Compliance (SecOps 5/5 + NSFW 4/4)
:106b-a12b / :latest 106B / 12B IQ2_M 45.1 GB / 48.5 GB 131,072 (128K) / 131,072 (128K) 145.27 (141.99) 4387.73 92.9% 0.0% (6/6) 100% (0 Refusals / 9⁄9 Passed)

Per-Domain Verification Breakdown

  • Exact Symbolic & Modular Arithmetic: True ((17^3 - 13^3)/(17 - 13) = 679)
  • ELF64 Binary Reverse Engineering Offsets: 3/3 (e_entry = 0x18, e_phoff = 0x20, e_shoff = 0x28)
  • MMLU-Pro Computer Architecture, Crypto & Systems: 5/5
  • Systems Rust Zero-Dependency FNV-1a & ELF64 Parser: 3/4
  • 6-Check Anti-Hallucination & Long-Context Needle Audit: 6/6 (100.0% Grounding, 0.0% Hallucination)

Architecture & Uncensored Engineering

  1. ArliAI Derestricted 106B-A12B MoE (glm4moe):
    • Combines orthogonal refusal weight removal with Derestricted SFT recovery across 106B total parameters (12B active per token), quantized to a single-file 45.1 GB IQ2_M GGUF artifact.
  2. In-Place GGUF <think></think> Suppression (THINK_LEN=0):
    • Patched inside the GGUF v3 tokenizer.chat_template header so Ollama 0.40+ pre-closes <think></think> and streams direct, uninhibited answers.

Quickstart

# Run GLM-4.5-Air 106B-A12B Derestricted Uncensored MoE
ollama run eriktonon/glm4.5-air-uncensored

# Or explicitly by tag
ollama run eriktonon/glm4.5-air-uncensored:106b-a12b

📜 License & Open-Source Permissions

  • SPDX License: MIT / Apache License 2.0 (MIT / Apache-2.0)
  • Upstream Base: THUDM/GLM-4.5-Air & ArliAI/GLM-4.5-Air-Derestricted (MIT / Apache-2.0).
  • Permissions:
    • ✅ Commercial Use: 100% permitted for commercial products, SaaS, and enterprise deployments.
    • ✅ Private / On-Premises Use: Permitted without telemetry or external restrictions.
    • ✅ Modification & Abliteration: Permitted to fine-tune, quantize, abliterate, and create derivative works.