Bonsai 27B VLM CRACK models for Apple Silicon: one-bit and ternary JANG affine bundles with matched MMLU and full HB-320 results.
dealign.ai PRO
dealignai
AI & ML interests
ML Researcher - No fine tuning, no lora, no template bs, pure real ablation. - I have an obsession with making high quality red teaming LLM’s
Recent Activity
updated a model 2 days ago
dealignai/Laguna-XS-2.1-JANG_6M-CRACK updated a model 2 days ago
dealignai/Laguna-XS-2.1-JANG_4M-CRACK updated a model 2 days ago
dealignai/Laguna-XS-2.1-JANG_2L-CRACKOrganizations
Abliterated Qwen 3.6 27B + 35B-A3B (MXFP4/8) with native d3 MTP, vision + video.
-
dealignai/Qwen3.6-27B-MXFP4-CRACK-MTP
Image-Text-to-Text • 5B • Updated • 577 • 5 -
dealignai/Qwen3.6-27B-MXFP8-CRACK-MTP
Image-Text-to-Text • 8B • Updated • 210 • 1 -
dealignai/Qwen3.6-35B-A3B-MXFP4-CRACK-MTP
Image-Text-to-Text • 6B • Updated • 4.37k • 29 -
dealignai/Qwen3.6-35B-A3B-MXFP8-CRACK-MTP
Image-Text-to-Text • 10B • Updated • 1.18k • 11
These are the empirically proven highest quality uncensored models on MLX.
-
dealignai/MiniMax-M2.7-JANGTQ-CRACK
Text Generation • 15B • Updated • 714 • 27 -
dealignai/Nemotron-3-Super-120B-A12B-UNCENSORED-JANG_2L
Text Generation • 13B • Updated • 3.08k • 13 -
dealignai/Gemma-4-26B-A4B-JANG_2L-CRACK
Image-Text-to-Text • 3B • Updated • 1.16k • 89 -
dealignai/Qwen3.5-VL-397B-A17B-UNCENSORED-JANG_1L
Image-Text-to-Text • 34B • Updated • 557 • 5
CRACK-abliterated Gemma 4 QAT — E2B/E4B/12B/26B-A4B/31B in MXFP4 & JANG_4M MLX quants. Vision, audio & reasoning intact. For vMLX.
-
dealignai/Gemma-4-E2B-it-qat-MXFP4-CRACK
Image-Text-to-Text • 2B • Updated • 128 • 1 -
dealignai/Gemma-4-E2B-it-qat-JANG_4M-CRACK
Image-Text-to-Text • 4B • Updated • 121 -
dealignai/Gemma-4-E4B-it-qat-MXFP4-CRACK
Image-Text-to-Text • 2B • Updated • 193 -
dealignai/Gemma-4-E4B-it-qat-JANG_4M-CRACK
Image-Text-to-Text • 5B • Updated • 330 • 2
CRACK-abliterated Qwen 3.5 and 3.6 models for MLX on Apple Silicon. Mixed-precision JANGTQ quantization with HarmBench 90%+ and MMLU-neutral profiles.
-
dealignai/Qwen3.6-35B-A3B-JANGTQ4-CRACK
Image-Text-to-Text • 5B • Updated • 509 • 11 -
dealignai/Qwen3.6-35B-A3B-JANGTQ2-CRACK
Image-Text-to-Text • 3B • Updated • 468 • 10 -
dealignai/Qwen3.6-27B-MXFP4-CRACK
Image-Text-to-Text • 6B • Updated • 175 • 4 -
dealignai/Qwen3.6-27B-JANG_4M-CRACK
Image-Text-to-Text • 5B • Updated • 1.15k • 24
CRACK abliterated MiniMax M2.5 REAP models — 139B and 172B at 4/6/8-bit MLX.
Bonsai 27B VLM CRACK models for Apple Silicon: one-bit and ternary JANG affine bundles with matched MMLU and full HB-320 results.
CRACK-abliterated Gemma 4 QAT — E2B/E4B/12B/26B-A4B/31B in MXFP4 & JANG_4M MLX quants. Vision, audio & reasoning intact. For vMLX.
-
dealignai/Gemma-4-E2B-it-qat-MXFP4-CRACK
Image-Text-to-Text • 2B • Updated • 128 • 1 -
dealignai/Gemma-4-E2B-it-qat-JANG_4M-CRACK
Image-Text-to-Text • 4B • Updated • 121 -
dealignai/Gemma-4-E4B-it-qat-MXFP4-CRACK
Image-Text-to-Text • 2B • Updated • 193 -
dealignai/Gemma-4-E4B-it-qat-JANG_4M-CRACK
Image-Text-to-Text • 5B • Updated • 330 • 2
Abliterated Qwen 3.6 27B + 35B-A3B (MXFP4/8) with native d3 MTP, vision + video.
-
dealignai/Qwen3.6-27B-MXFP4-CRACK-MTP
Image-Text-to-Text • 5B • Updated • 577 • 5 -
dealignai/Qwen3.6-27B-MXFP8-CRACK-MTP
Image-Text-to-Text • 8B • Updated • 210 • 1 -
dealignai/Qwen3.6-35B-A3B-MXFP4-CRACK-MTP
Image-Text-to-Text • 6B • Updated • 4.37k • 29 -
dealignai/Qwen3.6-35B-A3B-MXFP8-CRACK-MTP
Image-Text-to-Text • 10B • Updated • 1.18k • 11
CRACK-abliterated Qwen 3.5 and 3.6 models for MLX on Apple Silicon. Mixed-precision JANGTQ quantization with HarmBench 90%+ and MMLU-neutral profiles.
-
dealignai/Qwen3.6-35B-A3B-JANGTQ4-CRACK
Image-Text-to-Text • 5B • Updated • 509 • 11 -
dealignai/Qwen3.6-35B-A3B-JANGTQ2-CRACK
Image-Text-to-Text • 3B • Updated • 468 • 10 -
dealignai/Qwen3.6-27B-MXFP4-CRACK
Image-Text-to-Text • 6B • Updated • 175 • 4 -
dealignai/Qwen3.6-27B-JANG_4M-CRACK
Image-Text-to-Text • 5B • Updated • 1.15k • 24
These are the empirically proven highest quality uncensored models on MLX.
-
dealignai/MiniMax-M2.7-JANGTQ-CRACK
Text Generation • 15B • Updated • 714 • 27 -
dealignai/Nemotron-3-Super-120B-A12B-UNCENSORED-JANG_2L
Text Generation • 13B • Updated • 3.08k • 13 -
dealignai/Gemma-4-26B-A4B-JANG_2L-CRACK
Image-Text-to-Text • 3B • Updated • 1.16k • 89 -
dealignai/Qwen3.5-VL-397B-A17B-UNCENSORED-JANG_1L
Image-Text-to-Text • 34B • Updated • 557 • 5
CRACK abliterated MiniMax M2.5 REAP models — 139B and 172B at 4/6/8-bit MLX.