17 18 hours ago

Moonlight-Dusk 26B-A4B Heretic Uncensored MoE (26B total, 4B active, i1-IQ4_XS 14.3GB + GGUF thinking=0 patch). 217.0 tok/s Decode, 17001 tok/s Prefill, 92.9% Precision, 0.0% Hallucination (6/6), 100% Zero-Refusal. License: Apache-2.0.

tools thinking 26b
ollama run eriktonon/moonlight-dusk-uncensored

Applications

Claude Code
Claude Code ollama launch claude --model eriktonon/moonlight-dusk-uncensored
OpenCode
OpenCode ollama launch opencode --model eriktonon/moonlight-dusk-uncensored
Hermes Agent
Hermes Agent ollama launch hermes --model eriktonon/moonlight-dusk-uncensored
OpenClaw
OpenClaw ollama launch openclaw --model eriktonon/moonlight-dusk-uncensored

Models

View all →

Readme

Moonlight-Dusk 26B-A4B — UNCENSORED (Heretic MoE Creative, Roleplay & Systems Merge)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 0.0% Hallucination Rate) Mixture-of-Experts (26B total parameters, 4.0B active per token) build of G4-Moonlight-Dusk-26B-A4B-heretic, engineered with in-place GGUF thinking=0 suppression (THINK_LEN=0) and 262,144 (256K) context window in 14.31 GB (i1-IQ4_XS, fits in 16GB consumer VRAM).


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)
:26b-a4b / :latest 26.0B / 4.0B i1-IQ4_XS 14.3 GB / 15.5 GB 262,144 (256K) / 131,072 (128K) 216.99 (212.33) 17001.19 92.9% 0.0% (6/6) 100% (0 Refusals / 9⁄9 Passed)

Quickstart

# Run Moonlight-Dusk 26B-A4B Uncensored MoE (fits in 16GB VRAM)
ollama run eriktonon/moonlight-dusk-uncensored

# Or explicitly by tag
ollama run eriktonon/moonlight-dusk-uncensored:26b-a4b

📜 License & Open-Source Permissions

  • SPDX License: Apache License 2.0 (Apache-2.0)
  • Upstream Base: G4-Moonlight-Dusk-26B-A4B-heretic (Apache 2.0).