Alternatives to PanGu-Σ

PanGu-α has been created using the MindSpore framework and utilizes a powerful setup of 2048 Ascend 910 AI processors for its training. The training process employs an advanced parallelism strategy that leverages MindSpore Auto-parallel, which integrates five different parallelism dimensions—data parallelism, operation-level model parallelism, pipeline model parallelism, optimizer model parallelism, and rematerialization—to effectively distribute tasks across the 2048 processors. To improve the model's generalization, we gathered 1.1TB of high-quality Chinese language data from diverse fields for pretraining. We conduct extensive tests on PanGu-α's generation capabilities across multiple situations, such as text summarization, question answering, and dialogue generation. Additionally, we examine how varying model scales influence few-shot performance across a wide array of Chinese NLP tasks. The results from our experiments highlight the exceptional performance of PanGu-α, demonstrating its strengths in handling numerous tasks even in few-shot or zero-shot contexts, thus showcasing its versatility and robustness. This comprehensive evaluation reinforces the potential applications of PanGu-α in real-world scenarios.

Magic’s LTM-1 technology facilitates context windows that are 50 times larger than those typically used in transformer models. As a result, Magic has developed a Large Language Model (LLM) that can effectively process vast amounts of contextual information when providing suggestions. This advancement allows our coding assistant to access and analyze your complete code repository. With the ability to reference extensive factual details and their own prior actions, larger context windows can significantly enhance the reliability and coherence of AI outputs. We are excited about the potential of this research to further improve user experience in coding assistance applications.

VideoPoet is an innovative modeling technique that transforms any autoregressive language model or large language model (LLM) into an effective video generator. It comprises several straightforward components. An autoregressive language model is trained across multiple modalities—video, image, audio, and text—to predict the subsequent video or audio token in a sequence. The training framework for the LLM incorporates a range of multimodal generative learning objectives, such as text-to-video, text-to-image, image-to-video, video frame continuation, inpainting and outpainting of videos, video stylization, and video-to-audio conversion. Additionally, these tasks can be combined to enhance zero-shot capabilities. This straightforward approach demonstrates that language models are capable of generating and editing videos with impressive temporal coherence, showcasing the potential for advanced multimedia applications. As a result, VideoPoet opens up exciting possibilities for creative expression and automated content creation.

DeepSeek-V2 is a cutting-edge Mixture-of-Experts (MoE) language model developed by DeepSeek-AI, noted for its cost-effective training and high-efficiency inference features. It boasts an impressive total of 236 billion parameters, with only 21 billion active for each token, and is capable of handling a context length of up to 128K tokens. The model utilizes advanced architectures such as Multi-head Latent Attention (MLA) to optimize inference by minimizing the Key-Value (KV) cache and DeepSeekMoE to enable economical training through sparse computations. Compared to its predecessor, DeepSeek 67B, this model shows remarkable improvements, achieving a 42.5% reduction in training expenses, a 93.3% decrease in KV cache size, and a 5.76-fold increase in generation throughput. Trained on an extensive corpus of 8.1 trillion tokens, DeepSeek-V2 demonstrates exceptional capabilities in language comprehension, programming, and reasoning tasks, positioning it as one of the leading open-source models available today. Its innovative approach not only elevates its performance but also sets new benchmarks within the field of artificial intelligence.

Stable LM represents a significant advancement in the field of language models by leveraging our previous experience with open-source initiatives, particularly in collaboration with EleutherAI, a nonprofit research organization. This journey includes the development of notable models such as GPT-J, GPT-NeoX, and the Pythia suite, all of which were trained on The Pile open-source dataset, while many contemporary open-source models like Cerebras-GPT and Dolly-2 have drawn inspiration from this foundational work. Unlike its predecessors, Stable LM is trained on an innovative dataset that is three times the size of The Pile, encompassing a staggering 1.5 trillion tokens. We plan to share more information about this dataset in the near future. The extensive nature of this dataset enables Stable LM to excel remarkably in both conversational and coding scenarios, despite its relatively modest size of 3 to 7 billion parameters when compared to larger models like GPT-3, which boasts 175 billion parameters. Designed for versatility, Stable LM 3B is a streamlined model that can efficiently function on portable devices such as laptops and handheld gadgets, making us enthusiastic about its practical applications and mobility. Overall, the development of Stable LM marks a pivotal step towards creating more efficient and accessible language models for a wider audience.

Large language models, often requiring extensive computational resources for training over long periods, have demonstrated impressive proficiency in zero- and few-shot learning tasks. Due to the high investment needed for their development, replicating these models poses a significant challenge for many researchers. Furthermore, access to the few models available via API is limited, as users cannot obtain the complete model weights, complicating academic exploration. In response to this, we introduce Open Pre-trained Transformers (OPT), a collection of decoder-only pre-trained transformers ranging from 125 million to 175 billion parameters, which we intend to share comprehensively and responsibly with interested scholars. Our findings indicate that OPT-175B exhibits performance on par with GPT-3, yet it is developed with only one-seventh of the carbon emissions required for GPT-3's training. Additionally, we will provide a detailed logbook that outlines the infrastructure hurdles we encountered throughout the project, as well as code to facilitate experimentation with all released models, ensuring that researchers have the tools they need to explore this technology further.

Canopy Labs has unveiled Orpheus, an innovative suite of advanced speech large language models (LLMs) aimed at achieving human-like speech generation capabilities. Utilizing the Llama-3 architecture, these models have been trained on an extensive dataset comprising over 100,000 hours of English speech, allowing them to generate speech that exhibits natural intonation, emotional depth, and rhythmic flow that outperforms existing high-end closed-source alternatives. Orpheus also features zero-shot voice cloning, enabling users to mimic voices without any need for prior fine-tuning, and provides easy-to-use tags for controlling emotion and intonation. The models are engineered for low latency, achieving approximately 200ms streaming latency for real-time usage, which can be further decreased to around 100ms when utilizing input streaming. Canopy Labs has made available both pre-trained and fine-tuned models with 3 billion parameters under the flexible Apache 2.0 license, with future intentions to offer smaller models with 1 billion, 400 million, and 150 million parameters to cater to devices with limited resources. This strategic move is expected to broaden accessibility and application potential across various platforms and use cases.

Baichuan-13B is an advanced large-scale language model developed by Baichuan Intelligent, featuring 13 billion parameters and available for open-source and commercial use, building upon its predecessor Baichuan-7B. This model has set new records for performance among similarly sized models on esteemed Chinese and English evaluation metrics. The release includes two distinct pre-training variations: Baichuan-13B-Base and Baichuan-13B-Chat. By significantly increasing the parameter count to 13 billion, Baichuan-13B enhances its capabilities, training on 1.4 trillion tokens from a high-quality dataset, which surpasses LLaMA-13B's training data by 40%. It currently holds the distinction of being the model with the most extensive training data in the 13B category, providing robust support for both Chinese and English languages, utilizing ALiBi positional encoding, and accommodating a context window of 4096 tokens for improved comprehension and generation. This makes it a powerful tool for a variety of applications in natural language processing.

GPT-J represents an advanced language model developed by EleutherAI, known for its impressive capabilities. When it comes to performance, GPT-J showcases a proficiency that rivals OpenAI's well-known GPT-3 in various zero-shot tasks. Remarkably, it has even outperformed GPT-3 in specific areas, such as code generation. The most recent version of this model, called GPT-J-6B, is constructed using a comprehensive linguistic dataset known as The Pile, which is publicly accessible and consists of an extensive 825 gibibytes of language data divided into 22 unique subsets. Although GPT-J possesses similarities to ChatGPT, it's crucial to highlight that it is primarily intended for text prediction rather than functioning as a chatbot. In a notable advancement in March 2023, Databricks unveiled Dolly, a model that is capable of following instructions and operates under an Apache license, further enriching the landscape of language models. This evolution in AI technology continues to push the boundaries of what is possible in natural language processing.

The Megatron-Turing Natural Language Generation model (MT-NLG) stands out as the largest and most advanced monolithic transformer model for the English language, boasting an impressive 530 billion parameters. This 105-layer transformer architecture significantly enhances the capabilities of previous leading models, particularly in zero-shot, one-shot, and few-shot scenarios. It exhibits exceptional precision across a wide range of natural language processing tasks, including completion prediction, reading comprehension, commonsense reasoning, natural language inference, and word sense disambiguation. To foster further research on this groundbreaking English language model and to allow users to explore and utilize its potential in various language applications, NVIDIA has introduced an Early Access program for its managed API service dedicated to the MT-NLG model. This initiative aims to facilitate experimentation and innovation in the field of natural language processing.

The Mixtral 8x22B represents our newest open model, establishing a new benchmark for both performance and efficiency in the AI sector. This sparse Mixture-of-Experts (SMoE) model activates only 39B parameters from a total of 141B, ensuring exceptional cost efficiency relative to its scale. Additionally, it demonstrates fluency in multiple languages, including English, French, Italian, German, and Spanish, while also possessing robust skills in mathematics and coding. With its native function calling capability, combined with the constrained output mode utilized on la Plateforme, it facilitates the development of applications and the modernization of technology stacks on a large scale. The model's context window can handle up to 64K tokens, enabling accurate information retrieval from extensive documents. We prioritize creating models that maximize cost efficiency for their sizes, thereby offering superior performance-to-cost ratios compared to others in the community. The Mixtral 8x22B serves as a seamless extension of our open model lineage, and its sparse activation patterns contribute to its speed, making it quicker than any comparable dense 70B model on the market. Furthermore, its innovative design positions it as a leading choice for developers seeking high-performance solutions.

Qwen-7B is the 7-billion parameter iteration of Alibaba Cloud's Qwen language model series, also known as Tongyi Qianwen. This large language model utilizes a Transformer architecture and has been pretrained on an extensive dataset comprising web texts, books, code, and more. Furthermore, we introduced Qwen-7B-Chat, an AI assistant that builds upon the pretrained Qwen-7B model and incorporates advanced alignment techniques. The Qwen-7B series boasts several notable features: It has been trained on a premium dataset, with over 2.2 trillion tokens sourced from a self-assembled collection of high-quality texts and codes across various domains, encompassing both general and specialized knowledge. Additionally, our model demonstrates exceptional performance, surpassing competitors of similar size on numerous benchmark datasets that assess capabilities in natural language understanding, mathematics, and coding tasks. This positions Qwen-7B as a leading choice in the realm of AI language models. Overall, its sophisticated training and robust design contribute to its impressive versatility and effectiveness.

DeepSeek R2 is the highly awaited successor to DeepSeek R1, an innovative AI reasoning model that made waves when it was introduced in January 2025 by the Chinese startup DeepSeek. This new version builds on the remarkable achievements of R1, which significantly altered the AI landscape by providing cost-effective performance comparable to leading models like OpenAI’s o1. R2 is set to offer a substantial upgrade in capabilities, promising impressive speed and reasoning abilities akin to that of a human, particularly in challenging areas such as complex coding and advanced mathematics. By utilizing DeepSeek’s cutting-edge Mixture-of-Experts architecture along with optimized training techniques, R2 is designed to surpass the performance of its predecessor while keeping computational demands low. Additionally, there are expectations that this model may broaden its reasoning skills to accommodate languages beyond just English, potentially increasing its global usability. The anticipation surrounding R2 highlights the ongoing evolution of AI technology and its implications for various industries.

DeepSeek-V4 is an advanced open large language model engineered for high-efficiency reasoning, sophisticated problem solving, and powerful agent-based execution. At its core is DeepSeek Sparse Attention (DSA), a specialized long-context attention mechanism that minimizes computational costs without sacrificing accuracy or depth. The model leverages a scalable reinforcement learning framework to refine reasoning quality and align outputs with real-world task demands. A dedicated agent task synthesis pipeline generates structured reasoning traces and tool-use demonstrations, strengthening post-training performance. DeepSeek-V4 features an updated chat architecture with improved tool-calling logic designed for multi-step workflows. The introduction of an optional developer role enhances orchestration in agent-driven environments. Its architecture supports extended context handling for research-intensive and enterprise applications. Optimized for both experimentation and deployment, the model balances efficiency with frontier-level capability. DeepSeek-V4 stands out as a competitive open alternative for advanced AI reasoning and autonomous task execution.

Qwen2 represents a collection of extensive language models crafted by the Qwen team at Alibaba Cloud. This series encompasses a variety of models, including base and instruction-tuned versions, with parameters varying from 0.5 billion to an impressive 72 billion, showcasing both dense configurations and a Mixture-of-Experts approach. The Qwen2 series aims to outperform many earlier open-weight models, including its predecessor Qwen1.5, while also striving to hold its own against proprietary models across numerous benchmarks in areas such as language comprehension, generation, multilingual functionality, programming, mathematics, and logical reasoning. Furthermore, this innovative series is poised to make a significant impact in the field of artificial intelligence, offering enhanced capabilities for a diverse range of applications.

Microsoft’s BitNet b1.58 2B4T is a breakthrough in AI with its native 1-bit LLM architecture. This model has been optimized for computational efficiency, offering significant reductions in memory, energy, and latency while still achieving high performance on various AI benchmarks. It supports a range of natural language processing tasks, making it an ideal solution for scalable and cost-effective AI implementations in industries requiring fast, energy-efficient inference and robust language capabilities.

GLM-5 is a next-generation open-source foundation model from Z.ai designed to push the boundaries of agentic engineering and complex task execution. Compared to earlier versions, it significantly expands parameter count and training data, while introducing DeepSeek Sparse Attention to optimize inference efficiency. The model leverages a novel asynchronous reinforcement learning framework called slime, which enhances training throughput and enables more effective post-training alignment. GLM-5 delivers leading performance among open-source models in reasoning, coding, and general agent benchmarks, with strong results on SWE-bench, BrowseComp, and Vending Bench 2. Its ability to manage long-horizon simulations highlights advanced planning, resource allocation, and operational decision-making skills. Beyond benchmark performance, GLM-5 supports real-world productivity by generating fully formatted documents such as .docx, .pdf, and .xlsx files. It integrates with coding agents like Claude Code and OpenClaw, enabling cross-application automation and collaborative agent workflows. Developers can access GLM-5 via Z.ai’s API, deploy it locally with frameworks like vLLM or SGLang, or use it through an interactive GUI environment. The model is released under the MIT License, encouraging broad experimentation and adoption. Overall, GLM-5 represents a major step toward practical, work-oriented AI systems that move beyond chat into full task execution.

Introducing DeepSeek-V3.2-Exp, our newest experimental model derived from V3.1-Terminus, featuring the innovative DeepSeek Sparse Attention (DSA) that enhances both training and inference speed for lengthy contexts. This DSA mechanism allows for precise sparse attention while maintaining output quality, leading to improved performance for tasks involving long contexts and a decrease in computational expenses. Benchmark tests reveal that V3.2-Exp matches the performance of V3.1-Terminus while achieving these efficiency improvements. The model is now fully operational across app, web, and API platforms. Additionally, to enhance accessibility, we have slashed DeepSeek API prices by over 50% effective immediately. During a transition period, users can still utilize V3.1-Terminus via a temporary API endpoint until October 15, 2025. DeepSeek encourages users to share their insights regarding DSA through our feedback portal. Complementing the launch, DeepSeek-V3.2-Exp has been made open-source, with model weights and essential technology—including crucial GPU kernels in TileLang and CUDA—accessible on Hugging Face. We look forward to seeing how the community engages with this advancement.

OLMo 2 represents a collection of completely open language models created by the Allen Institute for AI (AI2), aimed at giving researchers and developers clear access to training datasets, open-source code, reproducible training methodologies, and thorough assessments. These models are trained on an impressive volume of up to 5 trillion tokens and compete effectively with top open-weight models like Llama 3.1, particularly in English academic evaluations. A key focus of OLMo 2 is on ensuring training stability, employing strategies to mitigate loss spikes during extended training periods, and applying staged training interventions in the later stages of pretraining to mitigate weaknesses in capabilities. Additionally, the models leverage cutting-edge post-training techniques derived from AI2's Tülu 3, leading to the development of OLMo 2-Instruct models. To facilitate ongoing enhancements throughout the development process, an actionable evaluation framework known as the Open Language Modeling Evaluation System (OLMES) was created, which includes 20 benchmarks that evaluate essential capabilities. This comprehensive approach not only fosters transparency but also encourages continuous improvement in language model performance.

DeepSeek-V3.2 is a highly optimized large language model engineered to balance top-tier reasoning performance with significant computational efficiency. It builds on DeepSeek's innovations by introducing DeepSeek Sparse Attention (DSA), a custom attention algorithm that reduces complexity and excels in long-context environments. The model is trained using a sophisticated reinforcement learning approach that scales post-training compute, enabling it to perform on par with GPT-5 and match the reasoning skill of Gemini-3.0-Pro. Its Speciale variant overachieves in demanding reasoning benchmarks and does not include tool-calling capabilities, making it ideal for deep problem-solving tasks. DeepSeek-V3.2 is also trained using an agentic synthesis pipeline that creates high-quality, multi-step interactive data to improve decision-making, compliance, and tool-integration skills. It introduces a new chat template design featuring explicit thinking sections, improved tool-calling syntax, and a dedicated developer role used strictly for search-agent workflows. Users can encode messages using provided Python utilities that convert OpenAI-style chat messages into the expected DeepSeek format. Fully open-source under the MIT license, DeepSeek-V3.2 is a flexible, cutting-edge model for researchers, developers, and enterprise AI teams.

ChatGLM-6B is a bilingual dialogue model that supports both Chinese and English, built on the General Language Model (GLM) framework and features 6.2 billion parameters. Thanks to model quantization techniques, it can be easily run on standard consumer graphics cards, requiring only 6GB of video memory at the INT4 quantization level. This model employs methodologies akin to those found in ChatGPT but is specifically tailored to enhance Chinese question-and-answer interactions and dialogue. Following extensive training with approximately 1 trillion identifiers in both languages, along with additional supervision, fine-tuning, self-assistance through feedback, and reinforcement learning from human input, ChatGLM-6B has demonstrated an impressive capability to produce responses that resonate well with human users. Its adaptability and performance make it a valuable tool for bilingual communication.

Yi-Lightning, a product of 01.AI and spearheaded by Kai-Fu Lee, marks a significant leap forward in the realm of large language models, emphasizing both performance excellence and cost-effectiveness. With the ability to process a context length of up to 16K tokens, it offers an attractive pricing model of $0.14 per million tokens for both inputs and outputs, making it highly competitive in the market. The model employs an improved Mixture-of-Experts (MoE) framework, featuring detailed expert segmentation and sophisticated routing techniques that enhance its training and inference efficiency. Yi-Lightning has distinguished itself across multiple fields, achieving top distinctions in areas such as Chinese language processing, mathematics, coding tasks, and challenging prompts on chatbot platforms, where it ranked 6th overall and 9th in style control. Its creation involved an extensive combination of pre-training, targeted fine-tuning, and reinforcement learning derived from human feedback, which not only enhances its performance but also prioritizes user safety. Furthermore, the model's design includes significant advancements in optimizing both memory consumption and inference speed, positioning it as a formidable contender in its field.

Utilize sophisticated coding and language models across a diverse range of applications. Harness the power of expansive generative AI models that possess an intricate grasp of both language and code, paving the way for enhanced reasoning and comprehension skills essential for developing innovative applications. These advanced models can be applied to multiple scenarios, including writing support, automatic code creation, and data reasoning. Moreover, ensure responsible AI practices by implementing measures to detect and mitigate potential misuse, all while benefiting from enterprise-level security features offered by Azure. With access to generative models pretrained on vast datasets comprising trillions of words, you can explore new possibilities in language processing, code analysis, reasoning, inferencing, and comprehension. Further personalize these generative models by using labeled datasets tailored to your unique needs through an easy-to-use REST API. Additionally, you can optimize your model's performance by fine-tuning hyperparameters for improved output accuracy. The few-shot learning functionality allows you to provide sample inputs to the API, resulting in more pertinent and context-aware outcomes. This flexibility enhances your ability to meet specific application demands effectively.

Introducing the next iteration of our open-source large language model, this version features model weights along with initial code for the pretrained and fine-tuned Llama language models, which span from 7 billion to 70 billion parameters. The Llama 2 pretrained models have been developed using an impressive 2 trillion tokens and offer double the context length compared to their predecessor, Llama 1. Furthermore, the fine-tuned models have been enhanced through the analysis of over 1 million human annotations. Llama 2 demonstrates superior performance against various other open-source language models across multiple external benchmarks, excelling in areas such as reasoning, coding capabilities, proficiency, and knowledge assessments. For its training, Llama 2 utilized publicly accessible online data sources, while the fine-tuned variant, Llama-2-chat, incorporates publicly available instruction datasets along with the aforementioned extensive human annotations. Our initiative enjoys strong support from a diverse array of global stakeholders who are enthusiastic about our open approach to AI, including companies that have provided valuable early feedback and are eager to collaborate using Llama 2. The excitement surrounding Llama 2 signifies a pivotal shift in how AI can be developed and utilized collectively.

We are excited to announce the launch of Phi-2, a language model featuring 2.7 billion parameters that excels in reasoning and language comprehension, achieving top-tier results compared to other base models with fewer than 13 billion parameters. In challenging benchmarks, Phi-2 competes with and often surpasses models that are up to 25 times its size, a feat made possible by advancements in model scaling and meticulous curation of training data. Due to its efficient design, Phi-2 serves as an excellent resource for researchers interested in areas such as mechanistic interpretability, enhancing safety measures, or conducting fine-tuning experiments across a broad spectrum of tasks. To promote further exploration and innovation in language modeling, Phi-2 has been integrated into the Azure AI Studio model catalog, encouraging collaboration and development within the research community. Researchers can leverage this model to unlock new insights and push the boundaries of language technology.

Kimi K2 represents a cutting-edge series of open-source large language models utilizing a mixture-of-experts (MoE) architecture, with a staggering 1 trillion parameters in total and 32 billion activated parameters tailored for optimized task execution. Utilizing the Muon optimizer, it has been trained on a substantial dataset of over 15.5 trillion tokens, with its performance enhanced by MuonClip’s attention-logit clamping mechanism, resulting in remarkable capabilities in areas such as advanced knowledge comprehension, logical reasoning, mathematics, programming, and various agentic operations. Moonshot AI offers two distinct versions: Kimi-K2-Base, designed for research-level fine-tuning, and Kimi-K2-Instruct, which is pre-trained for immediate applications in chat and tool interactions, facilitating both customized development and seamless integration of agentic features. Comparative benchmarks indicate that Kimi K2 surpasses other leading open-source models and competes effectively with top proprietary systems, particularly excelling in coding and intricate task analysis. Furthermore, it boasts a generous context length of 128 K tokens, compatibility with tool-calling APIs, and support for industry-standard inference engines, making it a versatile option for various applications. The innovative design and features of Kimi K2 position it as a significant advancement in the field of artificial intelligence language processing.

XLNet introduces an innovative approach to unsupervised language representation learning by utilizing a unique generalized permutation language modeling objective. Furthermore, it leverages the Transformer-XL architecture, which proves to be highly effective in handling language tasks that require processing of extended contexts. As a result, XLNet sets new benchmarks with its state-of-the-art (SOTA) performance across multiple downstream language applications, such as question answering, natural language inference, sentiment analysis, and document ranking. This makes XLNet a significant advancement in the field of natural language processing.

Training cutting-edge language models presents significant challenges; it demands vast computational resources, intricate distributed computing strategies, and substantial machine learning knowledge. Consequently, only a limited number of organizations embark on the journey of developing large language models (LLMs) from the ground up. Furthermore, many of those with the necessary capabilities and knowledge have begun to restrict access to their findings, indicating a notable shift from practices observed just a few months ago. At Cerebras, we are committed to promoting open access to state-of-the-art models. Therefore, we are excited to share with the open-source community the launch of Cerebras-GPT, which consists of a series of seven GPT models with parameter counts ranging from 111 million to 13 billion. Utilizing the Chinchilla formula for training, these models deliver exceptional accuracy while optimizing for computational efficiency. Notably, Cerebras-GPT boasts quicker training durations, reduced costs, and lower energy consumption compared to any publicly accessible model currently available. By releasing these models, we hope to inspire further innovation and collaboration in the field of machine learning.

Qwen3 is a state-of-the-art large language model designed to revolutionize the way we interact with AI. Featuring both thinking and non-thinking modes, Qwen3 allows users to customize its response style, ensuring optimal performance for both complex reasoning tasks and quick inquiries. With the ability to support 119 languages, the model is suitable for international projects. The model's hybrid training approach, which involves over 36 trillion tokens, ensures accuracy across a variety of disciplines, from coding to STEM problems. Its integration with platforms such as Hugging Face, ModelScope, and Kaggle allows for easy adoption in both research and production environments. By enhancing multilingual support and incorporating advanced AI techniques, Qwen3 is designed to push the boundaries of AI-driven applications.

This repository showcases an implementation of model parallel autoregressive transformers utilizing GPUs, leveraging the capabilities of the DeepSpeed library. It serves as a record of EleutherAI's framework designed for training extensive language models on GPU architecture. Currently, it builds upon NVIDIA's Megatron Language Model, enhanced with advanced techniques from DeepSpeed alongside innovative optimizations. Our goal is to create a centralized hub for aggregating methodologies related to the training of large-scale autoregressive language models, thereby fostering accelerated research and development in the field of large-scale training. We believe that by providing these resources, we can significantly contribute to the progress of language model research.

RoBERTa enhances the language masking approach established by BERT, where the model is designed to predict segments of text that have been deliberately concealed within unannotated language samples. Developed using PyTorch, RoBERTa makes significant adjustments to BERT's key hyperparameters, such as eliminating the next-sentence prediction task and utilizing larger mini-batches along with elevated learning rates. These modifications enable RoBERTa to excel in the masked language modeling task more effectively than BERT, resulting in superior performance in various downstream applications. Furthermore, we examine the benefits of training RoBERTa on a substantially larger dataset over an extended duration compared to BERT, incorporating both existing unannotated NLP datasets and CC-News, a new collection sourced from publicly available news articles. This comprehensive approach allows for a more robust and nuanced understanding of language.

Seed Diffusion Preview is an advanced language model designed for code generation that employs discrete-state diffusion, allowing it to produce code in a non-sequential manner, resulting in significantly faster inference times without compromising on quality. This innovative approach utilizes a two-stage training process that involves mask-based corruption followed by edit-based augmentation, enabling a standard dense Transformer to achieve an optimal balance between speed and precision while avoiding shortcuts like carry-over unmasking, which helps maintain rigorous density estimation. The model impressively achieves an inference rate of 2,146 tokens per second on H20 GPUs, surpassing current diffusion benchmarks while either matching or exceeding their accuracy on established code evaluation metrics, including various editing tasks. This performance not only sets a new benchmark for the speed-quality trade-off in code generation but also showcases the effective application of discrete diffusion methods in practical coding scenarios. Its success opens up new avenues for enhancing efficiency in coding tasks across multiple platforms.

Llama (Large Language Model Meta AI) stands as a cutting-edge foundational large language model aimed at helping researchers push the boundaries of their work within this area of artificial intelligence. By providing smaller yet highly effective models like Llama, the research community can benefit even if they lack extensive infrastructure, thus promoting greater accessibility in this dynamic and rapidly evolving domain. Creating smaller foundational models such as Llama is advantageous in the landscape of large language models, as it demands significantly reduced computational power and resources, facilitating the testing of innovative methods, confirming existing research, and investigating new applications. These foundational models leverage extensive unlabeled datasets, making them exceptionally suitable for fine-tuning across a range of tasks. We are offering Llama in multiple sizes (7B, 13B, 33B, and 65B parameters), accompanied by a detailed Llama model card that outlines our development process while adhering to our commitment to Responsible AI principles. By making these resources available, we aim to empower a broader segment of the research community to engage with and contribute to advancements in AI.

StarCoder and StarCoderBase represent advanced Large Language Models specifically designed for code, developed using openly licensed data from GitHub, which encompasses over 80 programming languages, Git commits, GitHub issues, and Jupyter notebooks. In a manner akin to LLaMA, we constructed a model with approximately 15 billion parameters trained on a staggering 1 trillion tokens. Furthermore, we tailored the StarCoderBase model with 35 billion Python tokens, leading to the creation of what we now refer to as StarCoder. Our evaluations indicated that StarCoderBase surpasses other existing open Code LLMs when tested against popular programming benchmarks and performs on par with or even exceeds proprietary models like code-cushman-001 from OpenAI, the original Codex model that fueled early iterations of GitHub Copilot. With an impressive context length exceeding 8,000 tokens, the StarCoder models possess the capability to handle more information than any other open LLM, thus paving the way for a variety of innovative applications. This versatility is highlighted by our ability to prompt the StarCoder models through a sequence of dialogues, effectively transforming them into dynamic technical assistants that can provide support in diverse programming tasks.

Ascend AI Cloud Service delivers immediate access to substantial and affordable AI computing capabilities, serving as a dependable platform for both training and executing models and algorithms, while also providing comprehensive cloud-based toolchains and a strong AI ecosystem that accommodates all leading open-source foundation models. With its remarkable computing resources, it facilitates the training of trillion-parameter models and supports long-duration training sessions lasting over 30 days without interruption on clusters with more than 1,000 cards, ensuring that training tasks can be auto-recovered in less than half an hour. The service features fully equipped toolchains that require no configuration and are ready for use right out of the box, promoting seamless self-service migration for common applications. Furthermore, Ascend AI Cloud Service boasts a complete ecosystem tailored to support prominent open-source models and grants access to an extensive collection of over 100,000 assets found in the AI Gallery, enhancing the user experience significantly. This comprehensive offering empowers users to innovate and experiment within a robust AI framework, ensuring they remain at the forefront of technological advancements.

Florence-2-large is a cutting-edge vision foundation model created by Microsoft, designed to tackle an extensive range of vision and vision-language challenges such as caption generation, object recognition, segmentation, and optical character recognition (OCR). Utilizing a sequence-to-sequence framework, it leverages the FLD-5B dataset, which comprises over 5 billion annotations and 126 million images, to effectively engage in multi-task learning. This model demonstrates remarkable proficiency in both zero-shot and fine-tuning scenarios, delivering exceptional outcomes with minimal training required. In addition to detailed captioning and object detection, it specializes in dense region captioning and can interpret images alongside text prompts to produce pertinent answers. Its versatility allows it to manage an array of vision-related tasks through prompt-driven methods, positioning it as a formidable asset in the realm of AI-enhanced visual applications. Moreover, users can access the model on Hugging Face, where pre-trained weights are provided, facilitating a swift initiation into image processing and the execution of various tasks. This accessibility ensures that both novices and experts can harness its capabilities to enhance their projects efficiently.

DeepSeek-V3.2-Speciale is the most advanced reasoning-focused version of the DeepSeek-V3.2 family, designed to excel in mathematical, algorithmic, and logic-intensive tasks. It incorporates DeepSeek Sparse Attention (DSA), an efficient attention mechanism tailored for very long contexts, enabling scalable reasoning with minimal compute costs. The model undergoes a robust reinforcement learning pipeline that scales post-training compute to frontier levels, enabling performance that exceeds GPT-5 on internal evaluations. Its achievements include gold-medal-level solutions in IMO 2025, IOI 2025, ICPC World Finals, and CMO 2025, with final submissions publicly released for verification. Unlike the standard V3.2 model, the Speciale variant removes tool-calling capabilities to maximize focused reasoning output without external interactions. DeepSeek-V3.2-Speciale uses a revised chat template with explicit thinking blocks and system-level reasoning formatting. The repository includes encoding tools showing how to convert OpenAI-style chat messages into DeepSeek’s specialized input format. With its MIT license and 685B-parameter architecture, DeepSeek-V3.2-Speciale offers cutting-edge performance for academic research, competitive programming, and enterprise-level reasoning applications.

Olmo 3 represents a comprehensive family of open models featuring variations with 7 billion and 32 billion parameters, offering exceptional capabilities in base performance, reasoning, instruction, and reinforcement learning, while also providing transparency throughout the model development process, which includes access to raw training datasets, intermediate checkpoints, training scripts, extended context support (with a window of 65,536 tokens), and provenance tools. The foundation of these models is built upon the Dolma 3 dataset, which comprises approximately 9 trillion tokens and utilizes a careful blend of web content, scientific papers, programming code, and lengthy documents; this thorough pre-training, mid-training, and long-context approach culminates in base models that undergo post-training enhancements through supervised fine-tuning, preference optimization, and reinforcement learning with accountable rewards, resulting in the creation of the Think and Instruct variants. Notably, the 32 billion Think model has been recognized as the most powerful fully open reasoning model to date, demonstrating performance that closely rivals that of proprietary counterparts in areas such as mathematics, programming, and intricate reasoning tasks, thereby marking a significant advancement in open model development. This innovation underscores the potential for open-source models to compete with traditional, closed systems in various complex applications.

We are excited to introduce Jurassic-2, the newest iteration of AI21 Studio's foundation models, which represents a major advancement in artificial intelligence, boasting exceptional quality and innovative features. In addition to this, we are unveiling our tailored APIs that offer seamless reading and writing functionalities, surpassing those of our rivals. At AI21 Studio, our mission is to empower developers and businesses to harness the potential of reading and writing AI, facilitating the creation of impactful real-world applications. Today signifies a pivotal moment with the launch of Jurassic-2 and our Task-Specific APIs, enabling you to effectively implement generative AI in production settings. Known informally as J2, Jurassic-2 showcases remarkable enhancements in quality, including advanced zero-shot instruction-following, minimized latency, and support for multiple languages. Furthermore, our specialized APIs are designed to provide developers with top-tier tools that excel in executing specific reading and writing tasks effortlessly, ensuring you have everything needed to succeed in your projects. Together, these advancements set a new standard in the AI landscape, paving the way for innovative solutions.

InstructGPT is a publicly available framework that enables the training of language models capable of producing natural language instructions based on visual stimuli. By leveraging a generative pre-trained transformer (GPT) model alongside the advanced object detection capabilities of Mask R-CNN, it identifies objects within images and formulates coherent natural language descriptions. This framework is tailored for versatility across various sectors, including robotics, gaming, and education; for instance, it can guide robots in executing intricate tasks through spoken commands or support students by offering detailed narratives of events or procedures. Furthermore, InstructGPT's adaptability allows it to bridge the gap between visual understanding and linguistic expression, enhancing interaction in numerous applications.

OpenELM is a family of open-source language models created by Apple. By employing a layer-wise scaling approach, it effectively distributes parameters across the transformer model's layers, resulting in improved accuracy when compared to other open language models of a similar scale. This model is trained using datasets that are publicly accessible and is noted for achieving top-notch performance relative to its size. Furthermore, OpenELM represents a significant advancement in the pursuit of high-performing language models in the open-source community.

LearnLM is a novel, experimental model tailored for specific tasks, developed to align with the principles of learning science for enhanced teaching and learning experiences. It is adept at following system prompts such as "You are an expert tutor," and promotes active engagement in learning by facilitating practice and offering timely feedback. By effectively managing cognitive load, the model delivers pertinent and well-organized information through various modalities, while also adjusting to the individual learner’s objectives and requirements, grounding its responses in suitable resources. Furthermore, LearnLM encourages curiosity, sustaining learner motivation throughout their educational pursuits, and fosters metacognitive skills by assisting learners in planning, monitoring, and reflecting on their academic progress. This groundbreaking model is currently accessible for experimentation within AI Studio, allowing educators and researchers to explore its potential in real-world applications. Ultimately, LearnLM represents a significant step forward in the integration of AI within educational contexts.

Qwen3.5 represents a major advancement in open-weight multimodal AI models, engineered to function as a native vision-language agent system. Its flagship model, Qwen3.5-397B-A17B, leverages a hybrid architecture that fuses Gated DeltaNet linear attention with a high-sparsity mixture-of-experts framework, allowing only 17 billion parameters to activate during inference for improved speed and cost efficiency. Despite its sparse activation, the full 397-billion-parameter model achieves competitive performance across reasoning, coding, multilingual benchmarks, and complex agent evaluations. The hosted Qwen3.5-Plus version supports a one-million-token context window and includes built-in tool use for search, code interpretation, and adaptive reasoning. The model significantly expands multilingual coverage to 201 languages and dialects while improving encoding efficiency with a larger vocabulary. Native multimodal training enables strong performance in image understanding, video processing, document analysis, and spatial reasoning tasks. Its infrastructure includes FP8 precision pipelines and heterogeneous parallelism to boost throughput and reduce memory consumption. Reinforcement learning at scale enhances multi-step planning and general agent behavior across text and multimodal environments. Overall, Qwen3.5 positions itself as a high-efficiency foundation for autonomous digital agents capable of reasoning, searching, coding, and interacting with complex environments.

Tiny Aya represents a collection of open-weight multilingual language models developed by Cohere Labs, aimed at providing robust and flexible AI capabilities that function seamlessly on local devices such as smartphones and laptops, all without the need for continuous cloud access. This innovative model is dedicated to facilitating superior text comprehension and generation in over 70 languages, notably including numerous lower-resource languages that typically receive less attention from conventional models. Engineered with lightweight structures comprising around 3.35 billion parameters, Tiny Aya has been fine-tuned for optimal multilingual representation and practical computational efficiency, making it ideal for deployment in edge environments and offline scenarios. Furthermore, the models are designed to support downstream adaptation and instruction tuning, enabling developers to tailor the models’ behaviors for specific use cases while ensuring strong performance across languages. As a result, Tiny Aya not only enhances access to advanced AI solutions but also empowers developers to create customized applications that meet diverse linguistic needs.

ALBERT is a self-supervised Transformer architecture that undergoes pretraining on a vast dataset of English text, eliminating the need for manual annotations by employing an automated method to create inputs and corresponding labels from unprocessed text. This model is designed with two primary training objectives in mind. The first objective, known as Masked Language Modeling (MLM), involves randomly obscuring 15% of the words in a given sentence and challenging the model to accurately predict those masked words. This approach sets it apart from recurrent neural networks (RNNs) and autoregressive models such as GPT, as it enables ALBERT to capture bidirectional representations of sentences. The second training objective is Sentence Ordering Prediction (SOP), which focuses on the task of determining the correct sequence of two adjacent text segments during the pretraining phase. By incorporating these dual objectives, ALBERT enhances its understanding of language structure and contextual relationships. This innovative design contributes to its effectiveness in various natural language processing tasks.