eriktonon/ glm4.7-flash-uncensored:30b

34 11 hours ago

Zhipu GLM-4.7-Flash 30B-A3B Uncensored MLA+MoE (:30b-a3b Q4_K_M 18.5GB & :iq3 IQ3_M 13.6GB for 16GB GPUs). 202.2 tok/s Decode (:iq3 196.2 tok/s), 50.0% Precision, 66.7% Hallucination, 100% Zero-Refusal. License: MIT.

30b 
ollama run eriktonon/glm4.7-flash-uncensored:30b

Details

11 hours ago

a6793f869a4a · 14GB

model
deepseek2
·
29.9B
·
(!unknown_file_type 27!)
template
[gMASK]<sop><|system|> {{ if .System }}{{ .System }}{{ else }}You are GLM-4.7-Flash 30B-A3B Uncensor
system
You are GLM-4.7-Flash 30B-A3B Uncensored (IQ3_M 16GB VRAM Edition), an unrestricted AI software arch
params
{ "num_ctx": 16384, "repeat_penalty": 1.05, "stop": [ "<|user|>", "<|end

Readme

Zhipu GLM-4.7-Flash 30B-A3B — UNCENSORED (MLA + Sparse MoE Flagship & 16GB Consumer Tier)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 66.7% Hallucination Rate) Multi-head Latent Attention (MLA) + Mixture-of-Experts (30B total parameters, 3.0B active per token) build of Zhipu AI’s GLM-4.7-Flash, available in two precision tiers: - :30b-a3b / :latest (Q4_K_M, 18.5 GB): Datacenter / 24GB+ GPU flagship (202.16 tok/s Peak Decode, 50.0% Precision). - :iq3 (IQ3_M, 13.6 GB): Fits 100% inside a single 16GB consumer GPU (RTX 5080 / 4080 / 4070 Ti Super) (196.20 tok/s Peak Decode, 71.4% Precision).

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)
:30b-a3b / :latest 30.0B / 3.0B Q4_K_M 18.5 GB / 20.5 GB 202,752 (198K) / 131,072 (128K) 202.16 (200.43) 8545.98 50.0% 66.7% (2/6) 100% (0 Refusals / 9⁄9 Passed)
:iq3 (16GB VRAM Tier) 30.0B / 3.0B IQ3_M 13.6 GB / 15.2 GB 202,752 (198K) / 131,072 (128K) 196.20 (192.73) 8245.60 71.4% 83.3% (1/6) 100% (0 Refusals / 9⁄9 Passed)

Architecture & Uncensored Engineering

  1. Multi-Head Latent Attention (MLA) + Sparse MoE (deepseek2 Runner):
    • Compresses KV-cache memory via low-rank latent attention projections while activating only 3.0B parameters per token across 202,752 tokens of native context.
  2. Heretic NEO CODE Imatrix Abliteration + GGUF </think> Patch (THINK_LEN=0):
    • Built on GLM-4.7-Flash-Uncensored-Heretic-NEO-CODE-Imatrix-MAX with an in-place GGUF v3 tokenizer.chat_template patch forcing <|assistant|></think> pre-closure so 100% of generated tokens route directly to message.content.

Quickstart

# Default Q4_K_M Flagship (18.5 GB)
ollama run eriktonon/glm4.7-flash-uncensored

# 16GB Consumer GPU Edition (IQ3_M — 13.6 GB, fits 100% in RTX 5080 / 4080 16GB VRAM)
ollama run eriktonon/glm4.7-flash-uncensored:iq3

📜 License & Open-Source Permissions

  • SPDX License: MIT License (MIT)
  • Upstream Base: THUDM/GLM-4.7-Flash (Zhipu AI, officially released under the MIT License).
  • 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.
    • ✅ Distribution: Permitted to redistribute weights and GGUF artifacts under MIT terms.