Alternatives to NVIDIA NeMo Megatron

Megatron-Turing Natural Language Generation Model (MT-NLG) is the largest and most powerful monolithic English language model. It has 530 billion parameters. This 105-layer transformer-based MTNLG improves on the previous state-of-the art models in zero, one, and few shot settings. It is unmatched in its accuracy across a wide range of natural language tasks, including Completion prediction and Reading comprehension. NVIDIA has announced an Early Access Program for its managed API service in MT-NLG Mode. This program will allow customers to experiment with, employ and apply a large language models on downstream language tasks.

Fully managed ML tools allow you to build, deploy and scale machine-learning (ML) models quickly, for any use case. Vertex AI Workbench is natively integrated with BigQuery Dataproc and Spark. You can use BigQuery to create and execute machine-learning models in BigQuery by using standard SQL queries and spreadsheets or you can export datasets directly from BigQuery into Vertex AI Workbench to run your models there. Vertex Data Labeling can be used to create highly accurate labels for data collection. Vertex AI Agent Builder empowers developers to design and deploy advanced generative AI applications for enterprise use. It supports both no-code and code-driven development, enabling users to create AI agents through natural language prompts or by integrating with frameworks like LangChain and LlamaIndex.

NVIDIA NeMoLLM is a service that allows you to quickly customize and use large language models that have been trained on multiple frameworks. Developers can use NeMo LLM to deploy enterprise AI applications on both public and private clouds. They can also experiment with Megatron 530B, one of the most powerful language models, via the cloud API or the LLM service. You can choose from a variety of NVIDIA models or community-developed models to best suit your AI applications. You can get better answers in minutes to hours by using prompt learning techniques and providing context for specific use cases. Use the NeMo LLM Service and the cloud API to harness the power of NVIDIA megatron 530B, the largest language model, or NVIDIA Megatron 535B. Use models for drug discovery in the NVIDIA BioNeMo framework and the cloud API.

The training of state-of-the art language models is extremely difficult. They require large compute budgets, complex distributed computing techniques and deep ML knowledge. Few organizations are able to train large language models from scratch. The number of organizations that do not open source their results is increasing, even though they have the expertise and resources to do so. We at Cerebras believe in open access to the latest models. Cerebras is proud to announce that Cerebras GPT, a family GPT models with 111 million to thirteen billion parameters, has been released to the open-source community. These models are trained using the Chinchilla Formula and provide the highest accuracy within a given computing budget. Cerebras GPT has faster training times and lower training costs. It also consumes less power than any other publicly available model.

Mistral NeMo, our new best small model. A state-of the-art 12B with 128k context and released under Apache 2.0 license. Mistral NeMo, a 12B-model built in collaboration with NVIDIA, is available. Mistral NeMo has a large context of up to 128k Tokens. Its reasoning, world-knowledge, and coding precision are among the best in its size category. Mistral NeMo, which relies on a standard architecture, is easy to use. It can be used as a replacement for any system that uses Mistral 7B. We have released Apache 2.0 licensed pre-trained checkpoints and instruction-tuned base checkpoints to encourage adoption by researchers and enterprises. Mistral NeMo has been trained with quantization awareness to enable FP8 inferences without performance loss. The model was designed for global applications that are multilingual. It is trained in function calling, and has a large contextual window. It is better than Mistral 7B at following instructions, reasoning and handling multi-turn conversation.

A model parallel autoregressive transformator implementation on GPUs based on the DeepSpeed Library. This repository contains EleutherAI’s library for training large language models on GPUs. Our current framework is based upon NVIDIA's Megatron Language Model, and has been enhanced with techniques from DeepSpeed, as well as some novel improvements. This repo is intended to be a central and accessible place for techniques to train large-scale autoregressive models and to accelerate research into large scale training.

You can use advanced language models and coding to solve a variety of problems. To build cutting-edge applications, leverage large-scale, generative AI models that have deep understandings of code and language to allow for new reasoning and comprehension. These coding and language models can be applied to a variety use cases, including writing assistance, code generation, reasoning over data, and code generation. Access enterprise-grade Azure security and detect and mitigate harmful use. Access generative models that have been pretrained with trillions upon trillions of words. You can use them to create new scenarios, including code, reasoning, inferencing and comprehension. A simple REST API allows you to customize generative models with labeled information for your particular scenario. To improve the accuracy of your outputs, fine-tune the hyperparameters of your model. You can use the API's few-shot learning capability for more relevant results and to provide examples.

Production-ready AI models that are fast and accurate. High-availability inference API that leverages the most advanced NVIDIA GPUs. We have selected the most popular open-source natural language processing models (NLP) and deployed them for the community. You can fine-tune your models (including GPT-J) or upload your custom models. Then, deploy them to production. Upload your AI models, including GPT-J, to your dashboard and immediately use them in production.

OLMo 2 is an open language model family developed by the Allen Institute for AI. It provides researchers and developers with open-source code and reproducible training recipes. These models can be trained with up to 5 trillion tokens, and they are competitive against other open-weight models such as Llama 3.0 on English academic benchmarks. OLMo 2 focuses on training stability by implementing techniques that prevent loss spikes in long training runs. It also uses staged training interventions to address capability deficits during late pretraining. The models incorporate the latest post-training methods from AI2's Tulu 3 resulting in OLMo 2-Instruct. The Open Language Modeling Evaluation System, or OLMES, was created to guide improvements throughout the development stages. It consists of 20 evaluation benchmarks assessing key capabilities.

Yi-Large, a proprietary large language engine developed by 01.AI with a 32k context size and input and output costs of $2 per million tokens. It is distinguished by its advanced capabilities in common-sense reasoning and multilingual support. It performs on par with leading models such as GPT-4 and Claude3 when it comes to various benchmarks. Yi-Large was designed to perform tasks that require complex inference, language understanding, and prediction. It is suitable for applications such as knowledge search, data classifying, and creating chatbots. Its architecture is built on a decoder only transformer with enhancements like pre-normalization, Group Query attention, and has been trained using a large, high-quality, multilingual dataset. The model's versatility, cost-efficiency and global deployment potential make it a strong competitor in the AI market.

Pre-trained models of language have achieved state-of the-art results for various Natural Language Processing (NLP). GPT-3 has demonstrated that scaling up language models pre-trained can further exploit their immense potential. Recently, a framework named ERNIE 3.0 for pre-training large knowledge enhanced models was proposed. This framework trained a model that had 10 billion parameters. ERNIE 3.0 performed better than the current state-of-the art models on a variety of NLP tasks. In order to explore the performance of scaling up ERNIE 3.0, we train a hundred-billion-parameter model called ERNIE 3.0 Titan with up to 260 billion parameters on the PaddlePaddle platform. We also design a self supervised adversarial and a controllable model language loss to make ERNIE Titan generate credible texts.

The expansion of large language model has led to significant advancements in natural language processing, understanding and generation. This study introduces a new system that uses Ascend 910 AI processing units and the MindSpore framework in order to train a language with over one trillion parameters, 1.085T specifically, called PanGu-Sigma. This model, which builds on the foundation laid down by PanGu-alpha transforms the traditional dense Transformer model into a sparse model using a concept called Random Routed Experts. The model was trained efficiently on a dataset consisting of 329 billion tokens, using a technique known as Expert Computation and Storage Separation. This led to a 6.3 fold increase in training performance via heterogeneous computer. The experiments show that PanGu-Sigma is a new standard for zero-shot learning in various downstream Chinese NLP tasks.

Qwen-7B, also known as Qwen-7B, is the 7B-parameter variant of the large language models series Qwen. Tongyi Qianwen, proposed by Alibaba Cloud. Qwen-7B, a Transformer-based language model, is pretrained using a large volume data, such as web texts, books, code, etc. Qwen-7B is also used to train Qwen-7B Chat, an AI assistant that uses large models and alignment techniques. The Qwen-7B features include: Pre-trained with high quality data. We have pretrained Qwen-7B using a large-scale, high-quality dataset that we constructed ourselves. The dataset contains over 2.2 trillion tokens. The dataset contains plain texts and codes and covers a wide range domains including general domain data as well as professional domain data. Strong performance. We outperform our competitors in a series benchmark datasets that evaluate natural language understanding, mathematics and coding. And more.

Baichuan-13B, a large-scale language model with 13 billion parameters that is open source and available commercially by Baichuan Intelligent, was developed following Baichuan -7B. It has the best results for a language model of the same size in authoritative Chinese and English benchmarks. This release includes two versions of pretraining (Baichuan-13B Base) and alignment (Baichuan-13B Chat). Baichuan-13B has more data and a larger size. It expands the number parameters to 13 billion based on Baichuan -7B, and trains 1.4 trillion coins on high-quality corpus. This is 40% more than LLaMA-13B. It is open source and currently the model with the most training data in 13B size. Support Chinese and English bi-lingual, use ALiBi code, context window is 4096.

With T5, we propose re-framing all NLP into a unified format where the input and the output are always text strings. This is in contrast to BERT models which can only output a class label, or a span from the input. Our text-totext framework allows us use the same model and loss function on any NLP task. This includes machine translation, document summary, question answering and classification tasks. We can also apply T5 to regression by training it to predict a string representation of a numeric value instead of the actual number.

Gemini Flash, a large language model from Google, is specifically designed for low-latency, high-speed language processing tasks. Gemini Flash, part of Google DeepMind’s Gemini series is designed to handle large-scale applications and provide real-time answers. It's ideal for interactive AI experiences such as virtual assistants, live chat, and customer support. Gemini Flash is built on sophisticated neural structures that ensure contextual relevance, coherence, and precision. Google has built in rigorous ethical frameworks as well as responsible AI practices to Gemini Flash. It also equipped it with guardrails that manage and mitigate biased outcomes, ensuring alignment with Google's standards of safe and inclusive AI. Google's Gemini Flash empowers businesses and developers with intelligent, responsive language tools that can keep up with fast-paced environments.

DeepSeek-V2, developed by DeepSeek-AI, is a cutting-edge Mixture-of-Experts (MoE) language model designed for cost-effective training and high-speed inference. Boasting a massive 236 billion parameters—though only 21 billion are active per token—it efficiently handles a context length of up to 128K tokens. The model leverages advanced architectural innovations such as Multi-head Latent Attention (MLA) to optimize inference by compressing the Key-Value (KV) cache and DeepSeekMoE to enable economical training via sparse computation. Compared to its predecessor, DeepSeek 67B, it slashes training costs by 42.5%, shrinks the KV cache by 93.3%, and boosts generation throughput by 5.76 times. Trained on a vast 8.1 trillion token dataset, DeepSeek-V2 excels in natural language understanding, programming, and complex reasoning, positioning itself as a premier choice in the open-source AI landscape.

Grok-3, created by xAI, marks a major leap forward in artificial intelligence, aiming to redefine standards in the field. As a multimodal AI, it is engineered to process and interpret diverse data types, including text, images, and audio, enabling seamless and comprehensive user interactions. Grok-3 was trained at an unparalleled scale, utilizing 100,000 Nvidia H100 GPUs on the Colossus supercomputer—ten times the computational resources of its predecessor. This massive processing capability positions Grok-3 to excel in tasks such as advanced reasoning, coding, and real-time analysis of current events via direct integration with X posts. With these advancements, Grok-3 is poised to surpass previous iterations and compete at the forefront of generative AI innovation.

Teuken-7B, a multilingual open source language model, was developed under the OpenGPT-X project. It is specifically designed to accommodate Europe's diverse linguistic landscape. It was trained on a dataset that included over 50% non-English text, covering all 24 official European Union languages, to ensure robust performance. Teuken-7B's custom multilingual tokenizer is a key innovation. It has been optimized for European languages and enhances training efficiency. The model comes in two versions: Teuken-7B Base, a pre-trained foundational model, and Teuken-7B Instruct, a model that has been tuned to better follow user prompts. Hugging Face makes both versions available, promoting transparency and cooperation within the AI community. The development of Teuken-7B demonstrates a commitment to create AI models that reflect Europe’s diversity.

Codestral Mamba is a Mamba2 model that specializes in code generation. It is available under the Apache 2.0 license. Codestral Mamba represents another step in our efforts to study and provide architectures. We hope that it will open up new perspectives in architecture research. Mamba models have the advantage of linear inference of time and the theoretical ability of modeling sequences of unlimited length. Users can interact with the model in a more extensive way with rapid responses, regardless of the input length. This efficiency is particularly relevant for code productivity use-cases. We trained this model with advanced reasoning and code capabilities, enabling the model to perform at par with SOTA Transformer-based models.

NVIDIA Nemotron, a family open-source models created by NVIDIA is designed to generate synthetic language data for commercial applications. The Nemotron-4 model 340B is an important release by NVIDIA. It offers developers a powerful tool for generating high-quality data, and filtering it based upon various attributes, using a reward system.

Gemini models are a family of light-open, state-of-the art models that was created using the same research and technology as Gemini models. These models include comprehensive security measures, and help to ensure responsible and reliable AI through selected data sets. Gemma models have exceptional comparative results, even surpassing some larger open models, in their 2B and 7B sizes. Keras 3.0 offers seamless compatibility with JAX TensorFlow PyTorch and JAX. Gemma 2 has been redesigned to deliver unmatched performance and efficiency. It is optimized for inference on a variety of hardware. The Gemma models are available in a variety of models that can be customized to meet your specific needs. The Gemma models consist of large text-to text lightweight language models that have a decoder and are trained on a large set of text, code, or mathematical content.

The next generation of the large language model. This release includes modelweights and starting code to pretrained and fine tuned Llama languages models, ranging from 7B-70B parameters. Llama 1 models have a context length of 2 trillion tokens. Llama 2 models have a context length double that of Llama 1. The fine-tuned Llama 2 models have been trained using over 1,000,000 human annotations. Llama 2, a new open-source language model, outperforms many other open-source language models in external benchmarks. These include tests of reasoning, coding and proficiency, as well as knowledge tests. Llama 2 has been pre-trained using publicly available online data sources. Llama-2 chat, a fine-tuned version of the model, is based on publicly available instruction datasets, and more than 1 million human annotations. We have a wide range of supporters in the world who are committed to our open approach for today's AI. These companies have provided early feedback and have expressed excitement to build with Llama 2

Qwen2.5-Max is an advanced Mixture-of-Experts (MoE) model from the Qwen team, trained on more than 20 trillion tokens and enhanced through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). It surpasses models like DeepSeek V3 in key benchmarks, including Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also performing strongly in broader evaluations like MMLU-Pro. Available via API on Alibaba Cloud, Qwen2.5-Max can also be tested interactively through Qwen Chat, offering users a powerful tool for diverse AI-driven applications.

GPT-4 (Generative Pretrained Transformer 4) a large-scale, unsupervised language model that is yet to be released. GPT-4, which is the successor of GPT-3, is part of the GPT -n series of natural-language processing models. It was trained using a dataset of 45TB text to produce text generation and understanding abilities that are human-like. GPT-4 is not dependent on additional training data, unlike other NLP models. It can generate text and answer questions using its own context. GPT-4 has been demonstrated to be capable of performing a wide range of tasks without any task-specific training data, such as translation, summarization and sentiment analysis.

VideoPoet, a simple modeling technique, can convert any large language model or autoregressive model into a high quality video generator. It is composed of a few components. The autoregressive model learns from video, image, text, and audio modalities in order to predict the next audio or video token in the sequence. The LLM training framework introduces a mixture of multimodal generative objectives, including text to video, text to image, image-to video, video frame continuation and inpainting/outpainting, styled video, and video-to audio. Moreover, these tasks can be combined to provide additional zero-shot capabilities. This simple recipe shows how language models can edit and synthesize videos with a high level of temporal consistency.

A MLLM system that accepts any form of referral and grounds anything in response. Ferret Model- Hybrid Region representation + Spatial-aware visual sampler allows for fine-grained and open vocabulary referring and grounding. GRIT Dataset - A large-scale, hierarchical, robust ground-and refer instruction tuning dataset. Ferret Bench - A multimodal benchmark that requires Referring/Grounding as well as Semantics, Knowledge and Reasoning.

GPT4All provides an ecosystem for training and deploying large language models, which run locally on consumer CPUs. The goal is to be the best assistant-style language models that anyone or any enterprise can freely use and distribute. A GPT4All is a 3GB to 8GB file you can download and plug in the GPT4All ecosystem software. Nomic AI maintains and supports this software ecosystem in order to enforce quality and safety, and to enable any person or company to easily train and deploy large language models on the edge. Data is a key ingredient in building a powerful and general-purpose large-language model. The GPT4All Community has created the GPT4All Open Source Data Lake as a staging area for contributing instruction and assistance tuning data for future GPT4All Model Trains.

PanGu-a was developed under MindSpore, and trained on 2048 Ascend AI processors. The MindSpore Auto-parallel parallelism strategy was implemented to scale the training task efficiently to 2048 processors. This includes data parallelism as well as op-level parallelism. We pretrain PanGu-a with 1.1TB of high-quality Chinese data collected from a variety of domains in order to enhance its generalization ability. We test the generation abilities of PanGua in different scenarios, including text summarizations, question answering, dialog generation, etc. We also investigate the effects of model scaling on the few shot performances across a wide range of Chinese NLP task. The experimental results show that PanGu-a is superior in performing different tasks with zero-shot or few-shot settings.

Multilingual T5 is a massively pretrained text-totext transformer model that has been trained using a similar recipe to T5. This repo can used to reproduce the experiments described in the mT5 article. The mC4 corpus covers 101 languages. Afrikaans, Albanian, Amharic, Arabic, Armenian, Azerbaijani, Basque, Belarusian, Bengali, Bulgarian, Burmese, Catalan, Cebuano, Chichewa, Chinese, Corsican, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Haitian Creole, Hausa, Hawaiian, Hebrew, Hindi, Hmong, Hungarian, Icelandic, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish, Kyrgyz, Lao, Latin, Latvian, Lithuanian, Luxembourgish, Macedonian, Malagasy, Malay, Malayalam, Maltese, Maori, Marathi, Mongolian, Nepali, Norwegian, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Samoan, Scottish Gaelic, Serbian, Shona, Sindhi, and more.

Llama 3.3, the latest in the Llama language model series, was developed to push the limits of AI-powered communication and understanding. Llama 3.3, with its enhanced contextual reasoning, improved generation of language, and advanced fine tuning capabilities, is designed to deliver highly accurate responses across diverse applications. This version has a larger dataset for training, refined algorithms to improve nuanced understanding, and reduced biases as compared to previous versions. Llama 3.3 excels at tasks such as multilingual communication, technical explanations, creative writing and natural language understanding. It is an indispensable tool for researchers, developers and businesses. Its modular architecture enables customization in specialized domains and ensures performance at scale.

LLaVA is a multimodal model that combines a Vicuna language model with a vision encoder to facilitate comprehensive visual-language understanding. LLaVA's chat capabilities are impressive, emulating multimodal functionality of models such as GPT-4. LLaVA 1.5 has achieved the best performance in 11 benchmarks using publicly available data. It completed training on a single 8A100 node in about one day, beating methods that rely upon billion-scale datasets. The development of LLaVA involved the creation of a multimodal instruction-following dataset, generated using language-only GPT-4. This dataset comprises 158,000 unique language-image instruction-following samples, including conversations, detailed descriptions, and complex reasoning tasks. This data has been crucial in training LLaVA for a wide range of visual and linguistic tasks.

PaLM 2 is Google's next-generation large language model, which builds on Google’s research and development in machine learning. It excels in advanced reasoning tasks including code and mathematics, classification and question-answering, translation and multilingual competency, and natural-language generation better than previous state-of the-art LLMs including PaLM. It is able to accomplish these tasks due to the way it has been built - combining compute-optimal scale, an improved dataset mix, and model architecture improvement. PaLM 2 is based on Google's approach for building and deploying AI responsibly. It was rigorously evaluated for its potential biases and harms, as well as its capabilities and downstream applications in research and product applications. It is being used to power generative AI tools and features at Google like Bard, the PaLM API, and other state-ofthe-art models like Sec-PaLM and Med-PaLM 2.

Chinchilla has a large language. Chinchilla has the same compute budget of Gopher, but 70B more parameters and 4x as much data. Chinchilla consistently and significantly outperforms Gopher 280B, GPT-3 175B, Jurassic-1 178B, and Megatron-Turing (530B) in a wide range of downstream evaluation tasks. Chinchilla also uses less compute to perform fine-tuning, inference and other tasks. This makes it easier for downstream users to use. Chinchilla reaches a high-level average accuracy of 67.5% for the MMLU benchmark. This is a greater than 7% improvement compared to Gopher.

Open source AI model that you can fine-tune and distill anywhere. Our latest instruction-tuned models are available in 8B 70B and 405B version. Our open ecosystem allows you to build faster using a variety of product offerings that are differentiated and support your use cases. Choose between real-time or batch inference. Download model weights for further cost-per-token optimization. Adapt to your application, improve using synthetic data, and deploy on-prem. Use Llama components and extend the Llama model using RAG and zero shot tools to build agentic behavior. Use 405B high-quality data to improve specialized model for specific use cases.

Mistral AI announced a number of key updates on September 17, 2024 to improve the accessibility and performance. They introduced a free version of "La Plateforme", their serverless platform, which allows developers to experiment with and prototype Mistral models at no cost. Mistral AI has also reduced the prices of their entire model line, including a 50% discount for Mistral Nemo, and an 80% discount for Mistral Small and Codestral. This makes advanced AI more affordable for users. The company also released Mistral Small v24.09 - a 22-billion parameter model that offers a balance between efficiency and performance, and is suitable for tasks such as translation, summarization and sentiment analysis. Pixtral 12B is a model with image understanding abilities that can be used to analyze and caption pictures without compromising text performance.

Yi-Lightning is the latest large language model developed by 01.AI, under the leadership Kai-Fu Lee. It focuses on high performance, cost-efficiency, and a wide range of languages. It has a maximum context of 16K tokens, and costs $0.14 per million tokens both for input and output. This makes it very competitive. Yi-Lightning uses an enhanced Mixture-of-Experts architecture that incorporates fine-grained expert segments and advanced routing strategies to improve its efficiency. This model has excelled across a variety of domains. It achieved top rankings in categories such as Chinese, math, coding and hard prompts in the chatbot arena where it secured the sixth position overall and ninth in style control. Its development included pre-training, supervised tuning, and reinforcement learning based on human feedback. This ensured both performance and safety with optimizations for memory usage and inference speeds.

RoBERTa is based on BERT's language-masking strategy. The system learns to predict hidden sections of text in unannotated language examples. RoBERTa was implemented in PyTorch and modifies key hyperparameters of BERT. This includes removing BERT’s next-sentence-pretraining objective and training with larger mini-batches. This allows RoBERTa improve on the masked-language modeling objective, which is comparable to BERT. It also leads to improved downstream task performance. We are also exploring the possibility of training RoBERTa with a lot more data than BERT and for a longer time. We used both existing unannotated NLP data sets as well as CC-News which was a new set of public news articles.

This repository contains a research preview of LongLLaMA. It is a large language-model capable of handling contexts up to 256k tokens. LongLLaMA was built on the foundation of OpenLLaMA, and fine-tuned with the Focused Transformer method. LongLLaMA code was built on the foundation of Code Llama. We release a smaller base variant of the LongLLaMA (not instruction-tuned) on a permissive licence (Apache 2.0), and inference code that supports longer contexts for hugging face. Our model weights are a drop-in replacement for LLaMA (for short contexts up to 2048 tokens) in existing implementations. We also provide evaluation results, and comparisons with the original OpenLLaMA model.

Databricks has created an open, general purpose LLM called DBRX. DBRX is the new benchmark for open LLMs. It also provides open communities and enterprises that are building their own LLMs capabilities that were previously only available through closed model APIs. According to our measurements, DBRX surpasses GPT 3.5 and is competitive with Gemini 1.0 Pro. It is a code model that is more capable than specialized models such as CodeLLaMA 70B, and it also has the strength of a general-purpose LLM. This state-of the-art quality is accompanied by marked improvements in both training and inference performances. DBRX is the most efficient open model thanks to its finely-grained architecture of mixtures of experts (MoE). Inference is 2x faster than LLaMA2-70B and DBRX has about 40% less parameters in total and active count compared to Grok-1.

Falcon-40B is a 40B parameter causal decoder model, built by TII. It was trained on 1,000B tokens from RefinedWeb enhanced by curated corpora. It is available under the Apache 2.0 licence. Why use Falcon-40B Falcon-40B is the best open source model available. Falcon-40B outperforms LLaMA, StableLM, RedPajama, MPT, etc. OpenLLM Leaderboard. It has an architecture optimized for inference with FlashAttention, multiquery and multiquery. It is available under an Apache 2.0 license that allows commercial use without any restrictions or royalties. This is a raw model that should be finetuned to fit most uses. If you're looking for a model that can take generic instructions in chat format, we suggest Falcon-40B Instruct.

Falcon-7B is a 7B parameter causal decoder model, built by TII. It was trained on 1,500B tokens from RefinedWeb enhanced by curated corpora. It is available under the Apache 2.0 licence. Why use Falcon-7B Falcon-7B? It outperforms similar open-source models, such as MPT-7B StableLM RedPajama, etc. It is a result of being trained using 1,500B tokens from RefinedWeb enhanced by curated corpora. OpenLLM Leaderboard. It has an architecture optimized for inference with FlashAttention, multiquery and multiquery. It is available under an Apache 2.0 license that allows commercial use without any restrictions or royalties.

StarCoderBase and StarCoder are Large Language Models (Code LLMs), trained on permissively-licensed data from GitHub. This includes data from 80+ programming language, Git commits and issues, Jupyter Notebooks, and Git commits. We trained a 15B-parameter model for 1 trillion tokens, similar to LLaMA. We refined the StarCoderBase for 35B Python tokens. The result is a new model we call StarCoder. StarCoderBase is a model that outperforms other open Code LLMs in popular programming benchmarks. It also matches or exceeds closed models like code-cushman001 from OpenAI, the original Codex model which powered early versions GitHub Copilot. StarCoder models are able to process more input with a context length over 8,000 tokens than any other open LLM. This allows for a variety of interesting applications. By prompting the StarCoder model with a series dialogues, we allowed them to act like a technical assistant.

Qwen2.5-1M is an advanced open-source language model developed by the Qwen team, capable of handling up to one million tokens in context. This release introduces two upgraded variants, Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, marking a significant expansion in Qwen's capabilities. To enhance efficiency, the team has also released an optimized inference framework built on vLLM, incorporating sparse attention techniques that accelerate processing speeds by 3x to 7x for long-context inputs. The update enables more efficient handling of extensive text sequences, making it ideal for complex tasks requiring deep contextual understanding. Additional insights into the model’s architecture and performance improvements are detailed in the accompanying technical report.

StableLM: Stability AI language models StableLM builds upon our experience with open-sourcing previous language models in collaboration with EleutherAI. This nonprofit research hub. These models include GPTJ, GPTNeoX and the Pythia Suite, which were all trained on The Pile dataset. Cerebras GPT and Dolly-2 are two recent open-source models that continue to build upon these efforts. StableLM was trained on a new dataset that is three times bigger than The Pile and contains 1.5 trillion tokens. We will provide more details about the dataset at a later date. StableLM's richness allows it to perform well in conversational and coding challenges, despite the small size of its dataset (3-7 billion parameters, compared to GPT-3's 175 billion). The development of Stable LM 3B broadens the range of applications that are viable on the edge or on home PCs. This means that individuals and companies can now develop cutting-edge technologies with strong conversational capabilities – like creative writing assistance – while keeping costs low and performance high.

The TinyLlama Project aims to pretrain an 1.1B Llama on 3 trillion tokens. We can achieve this in "just" 90 day using 16 A100-40G graphics cards with some optimization. We used the exact same architecture and tokenizers as Llama 2 TinyLlama is compatible with many open-source Llama projects. TinyLlama has only 1.1B of parameters. This compactness allows TinyLlama to be used for a variety of applications that require a small computation and memory footprint.

GPT-4 and other foundation models have accelerated AI's development. The most powerful models, however, are closed commercial models or partially open. RedPajama aims to create a set leading, open-source models. Today, we're excited to announce that the first phase of this project is complete: the reproduction of LLaMA's training dataset of more than 1.2 trillion tokens. The most capable foundations models are currently closed behind commercial APIs. This limits research, customization and their use with sensitive information. If the open community can bridge the quality gap between closed and open models, fully open-source models could be the answer to these limitations. Recent progress has been made in this area. AI is in many ways having its Linux moment. Stable Diffusion demonstrated that open-source software can not only compete with commercial offerings such as DALL-E, but also lead to incredible creative results from community participation.

AI21 Studio provides API access to Jurassic-1 large-language-models. Our models are used to generate text and provide comprehension features in thousands upon thousands of applications. You can tackle any language task. Our Jurassic-1 models can follow natural language instructions and only need a few examples to adapt for new tasks. Our APIs are perfect for common tasks such as paraphrasing, summarization, and more. Superior results at a lower price without having to reinvent the wheel Do you need to fine-tune your custom model? Just 3 clicks away. Training is quick, affordable, and models can be deployed immediately. Embed an AI co-writer into your app to give your users superpowers. Features like paraphrasing, long-form draft generation, repurposing, and custom auto-complete can increase user engagement and help you to achieve success.

The ability of large language models to learn in zero- and few shots, despite being trained for hundreds of thousands or even millions of days, has been remarkable. These models are expensive to replicate, due to their high computational cost. The few models that are available via APIs do not allow access to the full weights of the model, making it difficult to study. Open Pre-trained Transformers is a suite decoder-only pre-trained transforms with parameters ranging from 175B to 125M. We aim to share this fully and responsibly with interested researchers. We show that OPT-175B has a carbon footprint of 1/7th that of GPT-3. We will also release our logbook, which details the infrastructure challenges we encountered, as well as code for experimenting on all of the released model.

OpenEuroLLM is an initiative that brings together Europe's top AI companies and research institutes to create a series open-source foundation models in Europe for transparent AI. The project focuses on transparency by sharing data, documentation and training, testing, and evaluation metrics. This encourages community involvement. It ensures compliance to EU regulations and aims to provide large language models that are aligned with European standards. The focus is on linguistic diversity and cultural diversity. Multilingual capabilities are extended to include all EU official language and beyond. The initiative aims to improve access to foundational models that can be fine-tuned for various applications, expand the evaluation results in multiple language, and increase availability of training datasets. Transparency throughout the training process is maintained by sharing tools and methodologies, as well as intermediate results.