How pipenetwork/kimi-k3-mlx works
This repository is an Apple-Silicon MLX port of Kimi-K3, with a text model, a video-capable vision tower, and a multimodal wrapper that expands each image placeholder into its full feature block before text inference. A second product workflow streams the enormous source checkpoint into MLX-compatible artifacts, with optional REAP calibration for smaller tiers. The graph distinguishes documented runtime wiring from conversion and calibration; some internal model and processor details are described only in the README, so those links cite it.
Main entry points: kimi_k3_vl.py Multimodal model · convert.py Streaming converter · modeling_kimi_k3.py Reference multimodal model · language.py Text adapter
Components 20
MLX inference
- Multimodal model
kimi_k3_vl/kimi_k3_vl.py - Wrapper config
kimi_k3_vl/config.py - Text adapter
kimi_k3_vl/language.py - Text tower
kimi_k3.py - Image expansion
kimi_k3_vl/kimi_k3_vl.py
Vision processing
- Vision adapter
kimi_k3_vl/vision.py - MoonViT tower
kimi_k3_vision.py - Patch merger
kimi_k3_vision.py
Reference implementation
- Prompt processor
reference/kimi_k3_processor.py - Reference config
reference/configuration_kimi_k3.py - Reference text model
reference/modeling_kimi_linear.py - Reference multimodal model
reference/modeling_kimi_k3.py
Model artifacts
- Streaming converter
scripts/convert.py - Memory wiring
scripts/mlxmem.py
Tier calibration
- REAP calibration
scripts/reap_calibrate.py - Calibration builder
scripts/make_calib.py
Other components
- Prompting user
- Source checkpoint
- MLX model artifact
- Calibration data
Connections 17
- Prompting user → Prompt processor: provides prompt
README.md - Prompt processor → Multimodal model: formats image prompt
README.md - Multimodal model → Wrapper config: uses config
kimi_k3_vl/kimi_k3_vl.py - Multimodal model → Text adapter: dispatches text
kimi_k3_vl/kimi_k3_vl.py - Multimodal model → Vision adapter: dispatches images
kimi_k3_vl/kimi_k3_vl.py - Multimodal model → Image expansion: expands features
kimi_k3_vl/kimi_k3_vl.py - Text adapter → Text tower: wraps model
kimi_k3_vl/language.py - Vision adapter → MoonViT tower: wraps tower
kimi_k3_vl/vision.py - MoonViT tower → Patch merger: projects features
README.md - Source checkpoint → Streaming converter: streams weights
scripts/convert.py - Streaming converter → MLX model artifact: writes artifact
README.md - Calibration data → Calibration builder: provides samples
- Calibration builder → REAP calibration: prepares calibration
scripts/make_calib.py - REAP calibration → Streaming converter: supplies tier plan
scripts/reap_calibrate.py - MLX model artifact → Memory wiring: measures inference
- Reference multimodal model → Reference config: uses configuration
reference/modeling_kimi_k3.py - Reference multimodal model → Reference text model: uses text components
reference/modeling_kimi_k3.py