MoE Models

MoE Models SpikingBrain: 100x Faster LLM Inference via Spike Sparsity

SpikingBrain leverages brain-inspired spike sparsity and Mixture of Experts (MoE) to drastically reduce LLM latency. It achieves a 100x TTFT speedup on 4M token sequences and supports MetaX or NVIDIA GPUs.

MoE Models Dots.LLM1: 142B MoE Model Trained on 11.2T Real-World Tokens

Dots.LLM1 is a 142B total parameter MoE model with 14B activated. Trained on 11.2T tokens of real text, it rivals Qwen2.5-72B in performance at a lower computational cost.

Vision-Language-Action Models Mantis: A Smarter Vision-Language-Action Model for Robots

Mantis gives robots the precision to handle complex tasks using disentangled visual foresight, progressive training, and adaptive temporal ensembling.

Voice & Speech Tools Qwen3-ASR-Studio: Real-Time Voice Recognition with PiP Mode

Qwen3-ASR-Studio converts speech into text with high efficiency. Upload files, record live via a waveform interface, add context hints, and utilize PiP mode for global voice input. All data remains stored locally.

Music Players Infinite Radio: The AI DJ That Adapts Music Genres to Your Screen

Infinite Radio combines Magenta’s real-time music model with active screen analysis. Dynamically change genres based on your open applications or visual display content. Runs locally via Docker.

Knowledge Management MindForger Review: A Private Markdown IDE for Personal Knowledge Management

Stop losing ideas in scattered files. MindForger combines a powerful Markdown editor with a knowledge graph that mimics human memory. Open source, private, and cross-platform.

Video Tools HunyuanVideo-Avatar: Emotion-Controlled Multi-Person Video Generation

Tencent's HunyuanVideo-Avatar generates emotion-controlled multi-person dialogue videos. This guide covers its core modules and provides a complete installation walkthrough for NVIDIA GPUs.