2.4T-param MoE for Apple Silicon. REAP-pruned because no standard quant fits 550GB. Every build measured per-layer against bf16.
Pipenetwork PRO
pipenetwork
AI & ML interests
AI & ML
Recent Activity
updated a collection 5 days ago
Qwen3.8-2.4T-A95B MLX (REAP) updated a collection 5 days ago
Qwen3.8-2.4T-A95B MLX (REAP) updated a collection 5 days ago
Qwen3.8-2.4T-A95B MLX (REAP)Organizations
None yet
Apple Silicon (MLX) builds of MiniMax-H3, the 33B joint video+audio diffusion transformer. Code: github.com/PipeNetwork/minimax-h3-mlx
-
pipenetwork/MiniMax-H3-MLX-8bit
Image-Text-to-Video • 9B • Updated • 2.3k • 6 -
pipenetwork/MiniMax-H3-MLX-6bit
Image-Text-to-Video • 8B • Updated • 869 • 2 -
pipenetwork/MiniMax-H3-MLX-4bit
Image-Text-to-Video • 7B • Updated • 1.62k • 1 -
pipenetwork/MiniMax-H3-MLX-bf16
Image-Text-to-Video • 33B • Updated • 1.21k • 6
975B-A41B multimodal MoE on Apple Silicon. Text+image+audio. Ported from scratch: github.com/PipeNetwork/inkling-mlx
NVIDIA Nemotron TwoTower 30B-A3B diffusion LM in MLX for Apple Silicon: AR tower + two-tower diffusion, 4/6/8-bit + bf16.
-
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-4bit
Text Generation • 32B • Updated • 44 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-6bit
Text Generation • 32B • Updated • 85 • 1 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-8bit
Text Generation • 32B • Updated • 35 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-bf16
Text Generation • 32B • Updated • 82
MLX (Apple Silicon) conversions of deepreinforce-ai/Ornith-1.0-397B (Qwen3.5-MoE, text-only). 4/6/8-bit and bf16.
First MLX builds of zai-org/GLM-5.2 (glm_moe_dsa MoE): 4/5/6/8-bit + a 512GB-friendly mixed.
First MLX (Apple Silicon) quantizations of prefeitura-rio/Rio-3.1-Open-30B (Qwen3-MoE): 4/5/6/8-bit.
MLX (Apple Silicon) quantizations of google/gemma-4-31B-it: 4/5/6/8-bit. Text-only.
MLX (Apple Silicon) text-only conversions of MiniMax-M3 (427B MoE): 3-bit to 8-bit plus a mixed-precision build.
-
pipenetwork/MiniMax-M3-MLX-8bit
Text Generation • 426B • Updated • 309 • 1 -
pipenetwork/MiniMax-M3-MLX-6bit
Text Generation • 426B • Updated • 295 • 1 -
pipenetwork/MiniMax-M3-MLX-4bit
Text Generation • 426B • Updated • 262 -
pipenetwork/MiniMax-M3-MLX-mixed-3_6bit
Text Generation • 426B • Updated • 385 • 2
MLX quants of Microsoft's FrogBoss-32B & FrogMini-14B (Qwen3 debugging finetunes, SWE-bench ~45% pass@1) for Apple Silicon.
-
pipenetwork/FrogMini-14B-2510-MLX-4bit
Text Generation • 15B • Updated • 16 -
pipenetwork/FrogMini-14B-2510-MLX-8bit
Text Generation • 15B • Updated • 10 -
pipenetwork/FrogBoss-32B-2510-MLX-4bit
Text Generation • 33B • Updated • 7 -
pipenetwork/FrogBoss-32B-2510-MLX-8bit
Text Generation • 33B • Updated • 7
Apple Silicon (MLX) build of Muse-Glimmer-30B, plus the runtime no released mlx-vlm has: github.com/PipeNetwork/muse-glimmer-mlx
276B-A12B multimodal MoE on Apple Silicon. Same code path as Inkling 975B. REAP-25 is free: github.com/PipeNetwork/inkling-mlx
-
pipenetwork/Inkling-Small-MLX-8bit
Image-Text-to-Text • 74B • Updated • 186 -
pipenetwork/Inkling-Small-MLX-6bit
Image-Text-to-Text • 58B • Updated • 155 -
pipenetwork/Inkling-Small-MLX-4bit
Image-Text-to-Text • 41B • Updated • 425 • 1 -
pipenetwork/Inkling-Small-MLX-REAP25-4bit
Image-Text-to-Text • 31B • Updated • 379 • 1
304B MoE on Apple Silicon. deepseek_v4 is in no released runtime — ported from scratch. Take mixed-4_8bit. github.com/PipeNetwork/deepseek-v4-mlx
-
pipenetwork/DeepSeek-V4-Flash-MLX-8bit
Text Generation • 81B • Updated • 569 -
pipenetwork/DeepSeek-V4-Flash-MLX-6bit
Text Generation • 64B • Updated • 387 -
pipenetwork/DeepSeek-V4-Flash-MLX-mixed-4_8bit
Text Generation • 47B • Updated • 690 • 1 -
pipenetwork/DeepSeek-V4-Flash-MLX-4bit
Text Generation • 46B • Updated • 575
MLX (Apple Silicon) conversion of Qwen/Qwen3.6-35B-A3B in NVFP4 — the MLX analog of nvidia/Qwen3.6-35B-A3B-NVFP4. Text-only.
REAP expert-pruned GLM-5.2 (smaller/faster). REAP-25 near-lossless (+2.3% PPL); REAP-37 +7.3%; REAP-50 smallest but +37.5%.
First MLX (Apple Silicon) quantizations of inclusionAI VISTA-9B and VISTA-4B (qwen3_5).
MLX quantizations of google/gemma-4-26B-A4B-it (MoE): 5/6/8-bit (4-bit available from mlx-community). Text-only.
MLX build of Kimi-K2.7-Code. Base is natively 4-bit (int4 experts + bf16 rest); this keeps experts at 4-bit and lifts non-expert layers to 6-bit.
First working MLX builds of H Company's Holo-3.1 vision-language computer-use agents (Qwen3.5-VL). Vision-validated. Apache-2.0.
-
pipenetwork/Holo-3.1-4B-MLX-4bit
Image-Text-to-Text • 1.0B • Updated • 163 • 3 -
pipenetwork/Holo-3.1-4B-MLX-8bit
Image-Text-to-Text • 2B • Updated • 21 • 1 -
pipenetwork/Holo-3.1-9B-MLX-4bit
Image-Text-to-Text • 2B • Updated • 70 -
pipenetwork/Holo-3.1-9B-MLX-8bit
Image-Text-to-Text • 3B • Updated • 67 • 1
MLX quants of NVIDIA Nemotron-3 for Apple Silicon: Ultra 550B (4/5/6/8-bit) and dense Nano-4B (4/8-bit), converted with mlx-lm.
-
pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit
Text Generation • 549B • Updated • 128 -
pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-8bit
Text Generation • 549B • Updated • 82 -
pipenetwork/NVIDIA-Nemotron-3-Nano-4B-MLX-8bit
1B • Updated • 5 -
pipenetwork/NVIDIA-Nemotron-3-Nano-4B-MLX-4bit
0.6B • Updated • 2
2.4T-param MoE for Apple Silicon. REAP-pruned because no standard quant fits 550GB. Every build measured per-layer against bf16.
Apple Silicon (MLX) build of Muse-Glimmer-30B, plus the runtime no released mlx-vlm has: github.com/PipeNetwork/muse-glimmer-mlx
Apple Silicon (MLX) builds of MiniMax-H3, the 33B joint video+audio diffusion transformer. Code: github.com/PipeNetwork/minimax-h3-mlx
-
pipenetwork/MiniMax-H3-MLX-8bit
Image-Text-to-Video • 9B • Updated • 2.3k • 6 -
pipenetwork/MiniMax-H3-MLX-6bit
Image-Text-to-Video • 8B • Updated • 869 • 2 -
pipenetwork/MiniMax-H3-MLX-4bit
Image-Text-to-Video • 7B • Updated • 1.62k • 1 -
pipenetwork/MiniMax-H3-MLX-bf16
Image-Text-to-Video • 33B • Updated • 1.21k • 6
276B-A12B multimodal MoE on Apple Silicon. Same code path as Inkling 975B. REAP-25 is free: github.com/PipeNetwork/inkling-mlx
-
pipenetwork/Inkling-Small-MLX-8bit
Image-Text-to-Text • 74B • Updated • 186 -
pipenetwork/Inkling-Small-MLX-6bit
Image-Text-to-Text • 58B • Updated • 155 -
pipenetwork/Inkling-Small-MLX-4bit
Image-Text-to-Text • 41B • Updated • 425 • 1 -
pipenetwork/Inkling-Small-MLX-REAP25-4bit
Image-Text-to-Text • 31B • Updated • 379 • 1
975B-A41B multimodal MoE on Apple Silicon. Text+image+audio. Ported from scratch: github.com/PipeNetwork/inkling-mlx
304B MoE on Apple Silicon. deepseek_v4 is in no released runtime — ported from scratch. Take mixed-4_8bit. github.com/PipeNetwork/deepseek-v4-mlx
-
pipenetwork/DeepSeek-V4-Flash-MLX-8bit
Text Generation • 81B • Updated • 569 -
pipenetwork/DeepSeek-V4-Flash-MLX-6bit
Text Generation • 64B • Updated • 387 -
pipenetwork/DeepSeek-V4-Flash-MLX-mixed-4_8bit
Text Generation • 47B • Updated • 690 • 1 -
pipenetwork/DeepSeek-V4-Flash-MLX-4bit
Text Generation • 46B • Updated • 575
NVIDIA Nemotron TwoTower 30B-A3B diffusion LM in MLX for Apple Silicon: AR tower + two-tower diffusion, 4/6/8-bit + bf16.
-
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-4bit
Text Generation • 32B • Updated • 44 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-6bit
Text Generation • 32B • Updated • 85 • 1 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-8bit
Text Generation • 32B • Updated • 35 -
pipenetwork/Nemotron-3-Nano-30B-A3B-context-mlx-bf16
Text Generation • 32B • Updated • 82
MLX (Apple Silicon) conversion of Qwen/Qwen3.6-35B-A3B in NVFP4 — the MLX analog of nvidia/Qwen3.6-35B-A3B-NVFP4. Text-only.
MLX (Apple Silicon) conversions of deepreinforce-ai/Ornith-1.0-397B (Qwen3.5-MoE, text-only). 4/6/8-bit and bf16.
REAP expert-pruned GLM-5.2 (smaller/faster). REAP-25 near-lossless (+2.3% PPL); REAP-37 +7.3%; REAP-50 smallest but +37.5%.
First MLX builds of zai-org/GLM-5.2 (glm_moe_dsa MoE): 4/5/6/8-bit + a 512GB-friendly mixed.
First MLX (Apple Silicon) quantizations of inclusionAI VISTA-9B and VISTA-4B (qwen3_5).
First MLX (Apple Silicon) quantizations of prefeitura-rio/Rio-3.1-Open-30B (Qwen3-MoE): 4/5/6/8-bit.
MLX quantizations of google/gemma-4-26B-A4B-it (MoE): 5/6/8-bit (4-bit available from mlx-community). Text-only.
MLX (Apple Silicon) quantizations of google/gemma-4-31B-it: 4/5/6/8-bit. Text-only.
MLX build of Kimi-K2.7-Code. Base is natively 4-bit (int4 experts + bf16 rest); this keeps experts at 4-bit and lifts non-expert layers to 6-bit.
MLX (Apple Silicon) text-only conversions of MiniMax-M3 (427B MoE): 3-bit to 8-bit plus a mixed-precision build.
-
pipenetwork/MiniMax-M3-MLX-8bit
Text Generation • 426B • Updated • 309 • 1 -
pipenetwork/MiniMax-M3-MLX-6bit
Text Generation • 426B • Updated • 295 • 1 -
pipenetwork/MiniMax-M3-MLX-4bit
Text Generation • 426B • Updated • 262 -
pipenetwork/MiniMax-M3-MLX-mixed-3_6bit
Text Generation • 426B • Updated • 385 • 2
First working MLX builds of H Company's Holo-3.1 vision-language computer-use agents (Qwen3.5-VL). Vision-validated. Apache-2.0.
-
pipenetwork/Holo-3.1-4B-MLX-4bit
Image-Text-to-Text • 1.0B • Updated • 163 • 3 -
pipenetwork/Holo-3.1-4B-MLX-8bit
Image-Text-to-Text • 2B • Updated • 21 • 1 -
pipenetwork/Holo-3.1-9B-MLX-4bit
Image-Text-to-Text • 2B • Updated • 70 -
pipenetwork/Holo-3.1-9B-MLX-8bit
Image-Text-to-Text • 3B • Updated • 67 • 1
MLX quants of Microsoft's FrogBoss-32B & FrogMini-14B (Qwen3 debugging finetunes, SWE-bench ~45% pass@1) for Apple Silicon.
-
pipenetwork/FrogMini-14B-2510-MLX-4bit
Text Generation • 15B • Updated • 16 -
pipenetwork/FrogMini-14B-2510-MLX-8bit
Text Generation • 15B • Updated • 10 -
pipenetwork/FrogBoss-32B-2510-MLX-4bit
Text Generation • 33B • Updated • 7 -
pipenetwork/FrogBoss-32B-2510-MLX-8bit
Text Generation • 33B • Updated • 7
MLX quants of NVIDIA Nemotron-3 for Apple Silicon: Ultra 550B (4/5/6/8-bit) and dense Nano-4B (4/8-bit), converted with mlx-lm.
-
pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-4bit
Text Generation • 549B • Updated • 128 -
pipenetwork/NVIDIA-Nemotron-3-Ultra-550B-A55B-MLX-8bit
Text Generation • 549B • Updated • 82 -
pipenetwork/NVIDIA-Nemotron-3-Nano-4B-MLX-8bit
1B • Updated • 5 -
pipenetwork/NVIDIA-Nemotron-3-Nano-4B-MLX-4bit
0.6B • Updated • 2