PipeNetwork

MLX port of Qwen3.8-2.4T-A95B (2.4T params, 512-expert MoE) with streaming quantization, REAP expert pruning, and per-layer evaluation for a model too large to run

MLX (Apple Silicon) port of Muse-Glimmer-30B. Validated against transformers; also a runtime for the existing MLX conversions, which no released mlx-vlm can load.

MLX (Apple Silicon) port of MiniMax-H3 — 33B joint video+audio diffusion. Validated against the diffusers reference; AdaLN precompute drops 13B at inference.

MLX port of moonshotai/Kimi-K3 (2.78T multimodal MoE): streaming converter, REAP expert pruning, and per-language expert-overlap analysis

MLX (Apple Silicon) port of deepseek-ai/DeepSeek-V4-Flash-0731 (304B MoE). From-scratch implementation of Hyper-Connections, hash-routed experts, learned KV compression and sparse-attention indexing — none of which exist in any released runtime.

Run Thinking Machines Lab's Inkling (975B-A41B MoE, multimodal) on Apple Silicon with MLX — 4/6/8-bit.

MLX conversion pipeline for Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic (vision-enabled, 4/6/8bit/bf16)

MLX (Apple Silicon) port of poolside/Laguna-S-2.1 (118B-A8B MoE). Quants: pipenetwork/Laguna-S-2.1-MLX-*