Search Results for "online"

...FlashMLA, DeepEP, DeepGEMM, 3FS, etc.) that together form DeepSeek’s infrastructure stack. The repo's README describes the project as sharing “humble building blocks” of their online service—code that is documented, deployed, and battle-tested in production. The timing of its opening matches DeepSeek’s “Open-Source Week” campaign (starting around February 2025) when they gradually released internal infrastructure components publicly. It is licensed under CC0-1.0 (Creative Commons Zero) to maximize openness.

Qwen3-Omni is a natively end-to-end multilingual omni-modal foundation model that processes text, images, audio, and video and delivers real-time streaming responses in text and natural speech. It uses a Thinker-Talker architecture with a Mixture-of-Experts (MoE) design, early text-first pretraining, and mixed multimodal training to support strong performance across all modalities without sacrificing text or image quality. The model supports 119 text languages, 19 speech input languages, and...

DiffRhythm is an open-source, diffusion-based model designed to generate full-length songs. Focused on music creation, it combines advanced AI techniques to produce coherent and creative audio compositions. The model utilizes a latent diffusion architecture, making it capable of producing high-quality, long-form music. It can be accessed on Huggingface, where users can interact with a demo or download the model for further use. DiffRhythm offers tools for both training and inference, and its...

FLUX.1 Krea [dev] is an open-source 12-billion parameter image generation model developed collaboratively by Krea and Black Forest Labs, designed to deliver superior aesthetic control and high image quality. It is a rectified-flow model distilled from the original Krea 1, providing enhanced sampling efficiency through classifier-free guidance distillation. The model supports generation at resolutions between 1024 and 1280 pixels with recommended inference steps between 28 and 32 for optimal...

The Video PreTraining (VPT) repository provides code and model artifacts for a project where agents learn to act by watching human gameplay videos—specifically, gameplay of Minecraft—using behavioral cloning. The idea is to learn general priors of control from large-scale, unlabeled video data, and then optionally fine-tune those priors for more goal-directed behavior via environment interaction. The repository contains demonstration models of different widths, fine-tuned variants (e.g. for...

...It builds upon the Llama 3.1 architecture and incorporates fine-tuning techniques like SFT, DPO, and quantization-aware training for improved alignment, efficiency, and safety. The model supports eight primary languages (including English, Spanish, Hindi, and Thai) and was trained on a curated mix of publicly available online data, with a December 2023 knowledge cutoff. Llama-3.2-1B is lightweight enough for deployment on constrained devices like smartphones, using formats like SpinQuant and QLoRA to reduce model size and latency. Despite its small size, it performs competitively across benchmarks such as MMLU, ARC, and TLDR summarization. The model is distributed under the Llama 3.2 Community License, requiring attribution and adherence to Meta’s Acceptable Use Policy.