Rewritten 2026-08-30: router/mHC/KDA aux kept unquantized. 4bit=QUASAR-init (KL -8.5% vs RTN). ~29-33 tok/s on M3 Ultra. Receipts in each repo.
ALISVOLATPROPRIIS PRO
avlp12
AI & ML interests
MLX quantization, Apple Silicon inference, MoE models, local LLM deployment, medical AI, geopolitical early warning systems, humanoid robotics, semiconductor/memory markets, multi-agent systems
Recent Activity
updated a model about 1 hour ago
avlp12/GLM-5.3-Flash-Alis-MLX-4bit updated a model 1 day ago
avlp12/GLM-5.3-Flash-Alis-MLX-8bit updated a model 1 day ago
avlp12/GLM-5.3-Flash-Alis-MLX-6bitOrganizations
None yet
Receipt-sealed ALIS-DWQ MLX quants of Laguna S 2.1 (117.6B MoE) for Apple Silicon — highest-quality 4.89 bpw and capacity-optimal 3.31 bpw builds.
Receipt-sealed layer-local ALIS-DWQ MLX quants of Inkling 975B multimodal MoE, built on 2x512GB M3 Ultra. Quality 3.7 / capacity 2.7 bpw / Q6 soon.
-
thinkingmachines/Inkling
Image-Text-to-Text • 952B • Updated • 186k • • 1.76k -
avlp12/Inkling-975B-Alis-MLX-Dynamic-3.7bpw
Image-Text-to-Text • 947B • Updated • 178 • 2 -
avlp12/Inkling-975B-Alis-MLX-Dynamic-6.6bpw
Image-Text-to-Text • Updated • 34 -
avlp12/Inkling-975B-Alis-MLX-Dynamic-2.7bpw
Image-Text-to-Text • 947B • Updated • 46
Field guide to all Alis ports: frontier MoE quants, image DiTs and multimodal ports — MLX-native on Apple Silicon.
Civitai checkpoint converted for mflux — 8-bit and 4-bit builds of the same weights.
One base model, four packagings: MLX dynamic 2.6 bpw (text/VLM) and the Ultra GGUF variants.
-
avlp12/Qwen3.5-35B-A3B-Alis-MLX-Dynamic-2.6bpw
Text Generation • 35B • Updated • 138 • 1 -
avlp12/Qwen3.5-35B-A3B-Alis-MLX-Dynamic-2.6bpw-VLM
Image-Text-to-Text • 4B • Updated • 140 -
avlp12/Qwen3.5-35B-A3B-Alis-Ultra-GGUF
Text Generation • 35B • Updated • 2.48k • 2 -
avlp12/Qwen3.5-35B-A3B-Alis-Ultra-Slim-GGUF
Text Generation • 35B • Updated • 46 • 1
One base model (GLM-5.2 744B MoE), four MLX quant builds from 512 GB-class to the 256 GB floor — plus the GLM-5.1 benchmark baseline.
-
avlp12/GLM-5.2-Alis-MLX-Dynamic-4.5bpw
Text Generation • 743B • Updated • 622 • 1 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-3.5bpw
Text Generation • 753B • Updated • 1.14k • 5 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-2.56bpw
Text Generation • 73B • Updated • 811 • 2 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-2.3bpw
Text Generation • 743B • Updated • 346 • 1
Qwen3.8-27B for Apple silicon — vision tower (333 tensors, bf16) and vendor MTP head both preserved and verified. 8-bit / 6-bit / 4-bit AWQ.
World-first MLX ports + quants of Motif-3-Beta (314B MoE, KO/EN). 8-bit / 4.5bpw / 2.3bpw. Non-commercial research.
GPQA-verified mixed-precision MLX quants of Qwen3.6 for Apple Silicon — ThinkingCap 4.6bpw beats the official Q4_K_M at smaller size.
Traced MLX port of ByteDance Lance + the frozen PyTorch source mirror (verification reference).
Pure-MLX port of the 12.9B image DiT — 8-bit reference and the smaller mixed 4/8 build.
Same base model and 3.6 bpw quantization — text-only primary plus the vision add-on.
Rewritten 2026-08-30: router/mHC/KDA aux kept unquantized. 4bit=QUASAR-init (KL -8.5% vs RTN). ~29-33 tok/s on M3 Ultra. Receipts in each repo.
Qwen3.8-27B for Apple silicon — vision tower (333 tensors, bf16) and vendor MTP head both preserved and verified. 8-bit / 6-bit / 4-bit AWQ.
Receipt-sealed ALIS-DWQ MLX quants of Laguna S 2.1 (117.6B MoE) for Apple Silicon — highest-quality 4.89 bpw and capacity-optimal 3.31 bpw builds.
World-first MLX ports + quants of Motif-3-Beta (314B MoE, KO/EN). 8-bit / 4.5bpw / 2.3bpw. Non-commercial research.
Receipt-sealed layer-local ALIS-DWQ MLX quants of Inkling 975B multimodal MoE, built on 2x512GB M3 Ultra. Quality 3.7 / capacity 2.7 bpw / Q6 soon.
-
thinkingmachines/Inkling
Image-Text-to-Text • 952B • Updated • 186k • • 1.76k -
avlp12/Inkling-975B-Alis-MLX-Dynamic-3.7bpw
Image-Text-to-Text • 947B • Updated • 178 • 2 -
avlp12/Inkling-975B-Alis-MLX-Dynamic-6.6bpw
Image-Text-to-Text • Updated • 34 -
avlp12/Inkling-975B-Alis-MLX-Dynamic-2.7bpw
Image-Text-to-Text • 947B • Updated • 46
GPQA-verified mixed-precision MLX quants of Qwen3.6 for Apple Silicon — ThinkingCap 4.6bpw beats the official Q4_K_M at smaller size.
Field guide to all Alis ports: frontier MoE quants, image DiTs and multimodal ports — MLX-native on Apple Silicon.
Traced MLX port of ByteDance Lance + the frozen PyTorch source mirror (verification reference).
Civitai checkpoint converted for mflux — 8-bit and 4-bit builds of the same weights.
Pure-MLX port of the 12.9B image DiT — 8-bit reference and the smaller mixed 4/8 build.
One base model, four packagings: MLX dynamic 2.6 bpw (text/VLM) and the Ultra GGUF variants.
-
avlp12/Qwen3.5-35B-A3B-Alis-MLX-Dynamic-2.6bpw
Text Generation • 35B • Updated • 138 • 1 -
avlp12/Qwen3.5-35B-A3B-Alis-MLX-Dynamic-2.6bpw-VLM
Image-Text-to-Text • 4B • Updated • 140 -
avlp12/Qwen3.5-35B-A3B-Alis-Ultra-GGUF
Text Generation • 35B • Updated • 2.48k • 2 -
avlp12/Qwen3.5-35B-A3B-Alis-Ultra-Slim-GGUF
Text Generation • 35B • Updated • 46 • 1
Same base model and 3.6 bpw quantization — text-only primary plus the vision add-on.
One base model (GLM-5.2 744B MoE), four MLX quant builds from 512 GB-class to the 256 GB floor — plus the GLM-5.1 benchmark baseline.
-
avlp12/GLM-5.2-Alis-MLX-Dynamic-4.5bpw
Text Generation • 743B • Updated • 622 • 1 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-3.5bpw
Text Generation • 753B • Updated • 1.14k • 5 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-2.56bpw
Text Generation • 73B • Updated • 811 • 2 -
avlp12/GLM-5.2-Alis-MLX-Dynamic-2.3bpw
Text Generation • 743B • Updated • 346 • 1