Pipe Network PIPE: Kimi K3 MLX 端口发布
Pipe Network 于 7 月 28 日发布了 Kimi K3 的开源 MLX 移植版。通过流媒体转换器和 REAP 修剪，将模型大小从 1.6 TB 减少到约 350 GB，使其能够在 Mac Studio 上运行。
活动: 2026年7月28日 UTC
Pipe Network
@pipenetwork
@pipenetwork
Run Kimi K3 on a Mac Studio 🫰
K3 is 2.8T parameters and 1.6TB on disk, which makes it impossible to run on Apple Silicon.
Until now.
Our MLX port is now open source: https://github.com/PipeNetwork/kimi-k3-mlx
To accomplish this, we solved two things:
1. We wrote a streaming converter that walks one layer at a time, so that mlx_lm doesn't need to materialize the whole model.
2. REAP pruning sits on top and scores all 896 experts against a calibration corpus to keep only ones your workload needs.
That's what brings K3 down to 350GB and inside a Mac Studio.
K3 is 2.8T parameters and 1.6TB on disk, which makes it impossible to run on Apple Silicon.
Until now.
Our MLX port is now open source: https://github.com/PipeNetwork/kimi-k3-mlx
To accomplish this, we solved two things:
1. We wrote a streaming converter that walks one layer at a time, so that mlx_lm doesn't need to materialize the whole model.
2. REAP pruning sits on top and scores all 896 experts against a calibration corpus to keep only ones your workload needs.
That's what brings K3 down to 350GB and inside a Mac Studio.
Pipe Network
@pipenetwork
@pipenetwork
Get the full recipe, converter, calibration and pruning scripts here: https://t.co/q3jf1q3GZG
Quants published on Hugging Face:
1. REAP73: https://t.co/FvuDPBAL6B
2. REAP80: https://t.co/DQx7he4jex
3. REAP73 specialized in Chinese + code: https://t.co/woJkmKhu2B
Quants published on Hugging Face:
1. REAP73: https://t.co/FvuDPBAL6B
2. REAP80: https://t.co/DQx7he4jex
3. REAP73 specialized in Chinese + code: https://t.co/woJkmKhu2B
活动即将开始
0
D
0
H
0
M
0
S
Y
28 7月 16:38 (UTC)
✕
✕