# Large language models

Upstage AI leverages powerful large language models and document processing engines to transform workflows and improve efficiency for enterprises. Its main advantages include high accuracy, high performance, and customized solutions for various industries. It is positioned to empower leading enterprises and enhance work efficiency.

ComfyGen is an adaptive workflow system focused on text-to-image generation that automates and tailors effective workflows by learning from user prompts. The emergence of this technology marks a shift from using a single model to incorporating multiple specialized components in complex workflows aimed at enhancing image generation quality. A key advantage of ComfyGen is its ability to automatically adjust workflows based on user text prompts, making it especially valuable for users who need to generate images in specific styles or themes.

DataComp-LM (DCLM) is a comprehensive framework for building and training large language models (LLMs), providing standardized corpora, efficient pre-training recipes based on the open_lm framework, and over 50 evaluation methods. DCLM supports researchers in experimenting with different data set construction strategies at different computational scales, from 411M to 7B parameter models. DCLM significantly improves model performance through optimized dataset design and has already facilitated the creation of multiple high-quality datasets that outperform all open datasets at different scales.

Make-An-Audio 2 is a text-to-audio generation technology based on diffusion models, co-developed by researchers from Zhejiang University, ByteDance, and the Chinese University of Hong Kong. This technology utilizes pre-trained large language models (LLMs) to parse text, optimizing for semantic alignment and temporal consistency, thereby improving the quality of generated audio. It also incorporates a feed-forward Transformer-based diffusion denoiser to enhance performance in generating variable-length audio and bolster the extraction of temporal information. Furthermore, by leveraging LLMs to convert abundant audio label data into audio-text datasets, the issue of time data scarcity is addressed.

With BrainSoup, you can create custom AI agents to handle tasks and automate workflows through natural language. Enhance AI capabilities with your data while maintaining the highest levels of privacy and security. BrainSoup supports multiple large language models and semantic core technologies, making AI agents more powerful and personalized.

FP6-LLM is a new supporting solution for large language models. Through six-bit quantization (FP6), it effectively reduces the model size while maintaining model quality across various applications. We present TC-FPx, the first complete GPU kernel design that uniformly supports various quantization bit widths for floating-point weights. By integrating the TC-FPx kernel into existing inference systems, we provide a new end-to-end support for quantized LLM inference (called FP6-LLM), achieving a better balance between inference cost and model quality. Experiments demonstrate that FP6-LLM enables inference of LLaMA-70b using a single GPU, achieving normalized inference throughput 1.69x to 2.65x higher than the FP16 baseline.

StyleTTS 2 is a text-to-speech (TTS) model that utilizes large speech language models (SLMs) for style diffusion and adversarial training, achieving human-level TTS synthesis. It employs a diffusion model to model style as a latent stochastic variable, generating the most appropriate style for the given text without relying on voice references. Furthermore, we utilize large pre-trained SLMs (such as WavLM) as discriminators and incorporate our innovative differentiable duration modeling for end-to-end training, enhancing the naturalness of the synthesized speech. StyleTTS 2 surpasses human recordings on the single-speaker LJSpeech dataset and matches them on the multi-speaker VCTK dataset, garnering recognition from native English-speaking evaluators. Additionally, when trained on the LibriTTS dataset, our model outperforms prior publicly available zero-shot extension models. By demonstrating the potential of style diffusion and adversarial training with large SLMs, this work achieves human-level TTS synthesis on both single and multi-speaker datasets.