RepositoryPipeNetwork/kimi-k3-mlxTool
Kimi-K3 highlights the architectural and hardware limits of running frontier MoE models on Apple Silicon. It establishes key building blocks for MLX, including custom attention layers and streaming weight conversion for models that exceed multi-terabyte thresholds.
This project is aimed at ML engineers and researchers exploring novel MoE structures, such as LatentMoE and KDA, within the MLX ecosystem. The codebase addresses practical conversion bottlenecks through streaming pipelines. While the port features high technical fidelity—such as evaluating SiTU-GLU in fp32 to prevent rounding errors—its operational deployment remains restricted by current Apple Silicon hardware ceilings.
Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.
Similar in content, found by the AI search.
Link:Moonshot AI
Moonshot AI has released Kimi K3, an open-weight native multimodal agentic model featuring 2.8 trillion total parameters and a 1-million-token context window. The architecture uses a sparse MoE setup that activates 104 billion parameters per token.
Link:PocketAiHub
PocketAiHub has released an unofficial MLX derivative family of Qwen3.8-27B with suppressed refusal behaviour ('abliteration'). The consolidated repository provides quantized builds ranging from experimental 2-bit AWQ to full BF16 optimised for Apple Silicon.
Repository:drumih/turbo-fieldfare
TurboFieldfare is a dedicated Swift and Metal runtime designed to run Google's Gemma 4 26B-A4B on Apple Silicon Macs within an approximately 2 GB RAM budget. It avoids loading the full 14.3 GB model into memory by keeping only the 1.35 GB core and KV cache in RAM while streaming the required mixture-of-experts weights from SSD on demand.
6.6kstarsSwift
Repository:studio-dots-ai/dots3-note-prev
Dots Studio has released dots3-note preview, the first open-weight model in the dots3 family. The Mixture-of-Experts architecture features 280 billion total parameters with 16 billion activated parameters, multimodal inputs, and context support up to 512K tokens.
123stars
Link:drumih
TurboFieldfare is an open-source inference engine written in Swift and Metal that executes 4-bit Gemma 4 26B-A4B-IT on Apple M-series Macs using roughly 2 GB of RAM. While the model weights require 14 GB of disk space, the engine keeps only the shared model base and KV cache in memory and streams routed expert layers on demand directly from the SSD.
Repository:JustVugg/colibri
Colibrì is a lightweight inference engine written in pure C with zero dependencies, designed to run frontier Mixture-of-Experts (MoE) models ranging from 744B to 2.8T parameters on consumer and heterogeneous hardware. It treats VRAM, RAM, and storage as a unified memory hierarchy, streaming experts on demand.
37.7kstarsC
Parts of this content were created with the assistance of AI.
