Alternatives to Yi-Lightning

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.

Qwen2.5-Max is an advanced Mixture-of-Experts (MoE) model created by the Qwen team, which has been pretrained on an extensive dataset of over 20 trillion tokens and subsequently enhanced through methods like Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). Its performance in evaluations surpasses that of models such as DeepSeek V3 across various benchmarks, including Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also achieving strong results in other tests like MMLU-Pro. This model is available through an API on Alibaba Cloud, allowing users to easily integrate it into their applications, and it can also be interacted with on Qwen Chat for a hands-on experience. With its superior capabilities, Qwen2.5-Max represents a significant advancement in AI model technology.

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.

We are thrilled to present DBRX, a versatile open LLM developed by Databricks. This innovative model achieves unprecedented performance on a variety of standard benchmarks, setting a new benchmark for existing open LLMs. Additionally, it equips both the open-source community and enterprises crafting their own LLMs with features that were once exclusive to proprietary model APIs; our evaluations indicate that it outperforms GPT-3.5 and competes effectively with Gemini 1.0 Pro. Notably, it excels as a code model, outperforming specialized counterparts like CodeLLaMA-70B in programming tasks, while also demonstrating its prowess as a general-purpose LLM. The remarkable quality of DBRX is complemented by significant enhancements in both training and inference efficiency. Thanks to its advanced fine-grained mixture-of-experts (MoE) architecture, DBRX elevates the efficiency of open models to new heights. In terms of inference speed, it can be twice as fast as LLaMA2-70B, and its total and active parameter counts are approximately 40% of those in Grok-1, showcasing its compact design without compromising capability. This combination of speed and size makes DBRX a game-changer in the landscape of open AI models.

Kimi K2 Thinking is a sophisticated open-source reasoning model created by Moonshot AI, specifically tailored for intricate, multi-step workflows where it effectively combines chain-of-thought reasoning with tool utilization across numerous sequential tasks. Employing a cutting-edge mixture-of-experts architecture, the model encompasses a staggering total of 1 trillion parameters, although only around 32 billion parameters are utilized during each inference, which enhances efficiency while retaining significant capability. It boasts a context window that can accommodate up to 256,000 tokens, allowing it to process exceptionally long inputs and reasoning sequences without sacrificing coherence. Additionally, it features native INT4 quantization, which significantly cuts down inference latency and memory consumption without compromising performance. Designed with agentic workflows in mind, Kimi K2 Thinking is capable of autonomously invoking external tools, orchestrating sequential logic steps—often involving around 200-300 tool calls in a single chain—and ensuring consistent reasoning throughout the process. Its robust architecture makes it an ideal solution for complex reasoning tasks that require both depth and efficiency.

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.

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.

The MAI-1 Preview marks the debut of Microsoft AI's fully in-house developed foundation model, utilizing a mixture-of-experts architecture for streamlined performance. This model has undergone extensive training on around 15,000 NVIDIA H100 GPUs, equipping it to adeptly follow user instructions and produce relevant text responses for common inquiries, thus illustrating a prototype for future Copilot functionalities. Currently accessible for public testing on LMArena, MAI-1 Preview provides an initial look at the platform's direction, with plans to introduce select text-driven applications in Copilot over the next few weeks aimed at collecting user insights and enhancing its capabilities. Microsoft emphasizes its commitment to integrating its proprietary models, collaborations with partners, and advancements from the open-source sector to dynamically enhance user experiences through millions of distinct interactions every day. This innovative approach illustrates Microsoft's dedication to continuously evolving its AI offerings.

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.

Yi-Large is an innovative proprietary large language model created by 01.AI, featuring an impressive context length of 32k and a cost structure of $2 for each million tokens for both inputs and outputs. Renowned for its superior natural language processing abilities, common-sense reasoning, and support for multiple languages, it competes effectively with top models such as GPT-4 and Claude3 across various evaluations. This model is particularly adept at handling tasks that involve intricate inference, accurate prediction, and comprehensive language comprehension, making it ideal for applications such as knowledge retrieval, data categorization, and the development of conversational chatbots that mimic human interaction. Built on a decoder-only transformer architecture, Yi-Large incorporates advanced features like pre-normalization and Group Query Attention, and it has been trained on an extensive, high-quality multilingual dataset to enhance its performance. The model's flexibility and economical pricing position it as a formidable player in the artificial intelligence landscape, especially for businesses looking to implement AI technologies on a global scale. Additionally, its ability to adapt to a wide range of use cases underscores its potential to revolutionize how organizations leverage language models for various needs.

Z.ai has unveiled its latest flagship model, GLM-4.5, which boasts an impressive 355 billion total parameters (with 32 billion active) and is complemented by the GLM-4.5-Air variant, featuring 106 billion total parameters (12 billion active), designed to integrate sophisticated reasoning, coding, and agent-like functions into a single framework. This model can switch between a "thinking" mode for intricate, multi-step reasoning and tool usage and a "non-thinking" mode that facilitates rapid responses, accommodating a context length of up to 128K tokens and enabling native function invocation. Accessible through the Z.ai chat platform and API, and with open weights available on platforms like HuggingFace and ModelScope, GLM-4.5 is adept at processing a wide range of inputs for tasks such as general problem solving, common-sense reasoning, coding from the ground up or within existing frameworks, as well as managing comprehensive workflows like web browsing and slide generation. The architecture is underpinned by a Mixture-of-Experts design, featuring loss-free balance routing, grouped-query attention mechanisms, and an MTP layer that facilitates speculative decoding, ensuring it meets enterprise-level performance standards while remaining adaptable to various applications. As a result, GLM-4.5 sets a new benchmark for AI capabilities across numerous domains.

Teuken-7B is a multilingual language model that has been developed as part of the OpenGPT-X initiative, specifically tailored to meet the needs of Europe's varied linguistic environment. This model has been trained on a dataset where over half consists of non-English texts, covering all 24 official languages of the European Union, which ensures it performs well across these languages. A significant advancement in Teuken-7B is its unique multilingual tokenizer, which has been fine-tuned for European languages, leading to enhanced training efficiency and lower inference costs when compared to conventional monolingual tokenizers. Users can access two versions of the model: Teuken-7B-Base, which serves as the basic pre-trained version, and Teuken-7B-Instruct, which has received instruction tuning aimed at boosting its ability to respond to user requests. Both models are readily available on Hugging Face, fostering an environment of transparency and collaboration within the artificial intelligence community while also encouraging further innovation. The creation of Teuken-7B highlights a dedication to developing AI solutions that embrace and represent the rich diversity found across Europe.

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.

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.

Introducing an open-source AI model that can be fine-tuned, distilled, and deployed across various platforms. Our newest instruction-tuned model comes in three sizes: 8B, 70B, and 405B, giving you options to suit different needs. With our open ecosystem, you can expedite your development process using a diverse array of tailored product offerings designed to meet your specific requirements. You have the flexibility to select between real-time inference and batch inference services according to your project's demands. Additionally, you can download model weights to enhance cost efficiency per token while fine-tuning for your application. Improve performance further by utilizing synthetic data and seamlessly deploy your solutions on-premises or in the cloud. Take advantage of Llama system components and expand the model's capabilities through zero-shot tool usage and retrieval-augmented generation (RAG) to foster agentic behaviors. By utilizing 405B high-quality data, you can refine specialized models tailored to distinct use cases, ensuring optimal functionality for your applications. Ultimately, this empowers developers to create innovative solutions that are both efficient and effective.

Introducing Mistral NeMo, our latest and most advanced small model yet, featuring a cutting-edge 12 billion parameters and an expansive context length of 128,000 tokens, all released under the Apache 2.0 license. Developed in partnership with NVIDIA, Mistral NeMo excels in reasoning, world knowledge, and coding proficiency within its category. Its architecture adheres to industry standards, making it user-friendly and a seamless alternative for systems currently utilizing Mistral 7B. To facilitate widespread adoption among researchers and businesses, we have made available both pre-trained base and instruction-tuned checkpoints under the same Apache license. Notably, Mistral NeMo incorporates quantization awareness, allowing for FP8 inference without compromising performance. The model is also tailored for diverse global applications, adept in function calling and boasting a substantial context window. When compared to Mistral 7B, Mistral NeMo significantly outperforms in understanding and executing detailed instructions, showcasing enhanced reasoning skills and the ability to manage complex multi-turn conversations. Moreover, its design positions it as a strong contender for multi-lingual tasks, ensuring versatility across various use cases.

Ai2 OLMoE is a completely open-source mixture-of-experts language model that operates entirely on-device, ensuring that you can experiment with the model in a private and secure manner. This application is designed to assist researchers in advancing on-device intelligence and to allow developers to efficiently prototype innovative AI solutions without the need for cloud connectivity. OLMoE serves as a highly efficient variant within the Ai2 OLMo model family. Discover the capabilities of state-of-the-art local models in performing real-world tasks, investigate methods to enhance smaller AI models, and conduct local tests of your own models utilizing our open-source codebase. Furthermore, you can seamlessly integrate OLMoE into various iOS applications, as the app prioritizes user privacy and security by functioning entirely on-device. Users can also easily share the outcomes of their interactions with friends or colleagues. Importantly, both the OLMoE model and the application code are fully open source, offering a transparent and collaborative approach to AI development. By leveraging this model, developers can contribute to the growing field of on-device AI while maintaining high standards of user privacy.

DeepSeek-R1 is a cutting-edge open-source reasoning model created by DeepSeek, aimed at competing with OpenAI's Model o1. It is readily available through web, app, and API interfaces, showcasing its proficiency in challenging tasks such as mathematics and coding, and achieving impressive results on assessments like the American Invitational Mathematics Examination (AIME) and MATH. Utilizing a mixture of experts (MoE) architecture, this model boasts a remarkable total of 671 billion parameters, with 37 billion parameters activated for each token, which allows for both efficient and precise reasoning abilities. As a part of DeepSeek's dedication to the progression of artificial general intelligence (AGI), the model underscores the importance of open-source innovation in this field. Furthermore, its advanced capabilities may significantly impact how we approach complex problem-solving in various domains.

Qwen3-Max represents Alibaba's cutting-edge large language model, featuring a staggering trillion parameters aimed at enhancing capabilities in tasks that require agency, coding, reasoning, and managing lengthy contexts. This model is an evolution of the Qwen3 series, leveraging advancements in architecture, training methods, and inference techniques; it integrates both thinker and non-thinker modes, incorporates a unique “thinking budget” system, and allows for dynamic mode adjustments based on task complexity. Capable of handling exceptionally lengthy inputs, processing hundreds of thousands of tokens, it also supports tool invocation and demonstrates impressive results across various benchmarks, including coding, multi-step reasoning, and agent evaluations like Tau2-Bench. While the initial version prioritizes instruction adherence in a non-thinking mode, Alibaba is set to introduce reasoning functionalities that will facilitate autonomous agent operations in the future. In addition to its existing multilingual capabilities and extensive training on trillions of tokens, Qwen3-Max is accessible through API interfaces that align seamlessly with OpenAI-style functionalities, ensuring broad usability across applications. This comprehensive framework positions Qwen3-Max as a formidable player in the realm of advanced artificial intelligence language models.

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.

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.

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.

Gemini 2.5, developed by Google DeepMind, represents a breakthrough in AI with enhanced reasoning capabilities and native multimodality, allowing it to process long context windows of up to one million tokens. The family includes three variants: Pro for complex coding tasks, Flash for fast general use, and Flash-Lite for high-volume, cost-efficient workflows. Gemini 2.5 models improve accuracy by thinking through diverse strategies and provide developers with adaptive controls to optimize performance and resource use. The models handle multiple input types—text, images, video, audio, and PDFs—and offer powerful tool use like search and code execution. Gemini 2.5 achieves state-of-the-art results across coding, math, science, reasoning, and multilingual benchmarks, outperforming its predecessors. It is accessible through Google AI Studio, Gemini API, and Vertex AI platforms. Google emphasizes responsible AI development, prioritizing safety and security in all applications. Gemini 2.5 enables developers to build advanced interactive simulations, automated coding, and other innovative AI-driven solutions.

Tülu 3 is a cutting-edge language model created by the Allen Institute for AI (Ai2) that aims to improve proficiency in fields like knowledge, reasoning, mathematics, coding, and safety. It is based on the Llama 3 Base and undergoes a detailed four-stage post-training regimen: careful prompt curation and synthesis, supervised fine-tuning on a wide array of prompts and completions, preference tuning utilizing both off- and on-policy data, and a unique reinforcement learning strategy that enhances targeted skills through measurable rewards. Notably, this open-source model sets itself apart by ensuring complete transparency, offering access to its training data, code, and evaluation tools, thus bridging the performance divide between open and proprietary fine-tuning techniques. Performance assessments reveal that Tülu 3 surpasses other models with comparable sizes, like Llama 3.1-Instruct and Qwen2.5-Instruct, across an array of benchmarks, highlighting its effectiveness. The continuous development of Tülu 3 signifies the commitment to advancing AI capabilities while promoting an open and accessible approach to technology.

Qwen3-Coder is a versatile coding model that comes in various sizes, prominently featuring the 480B-parameter Mixture-of-Experts version with 35B active parameters, which naturally accommodates 256K-token contexts that can be extended to 1M tokens. This model achieves impressive performance that rivals Claude Sonnet 4, having undergone pre-training on 7.5 trillion tokens, with 70% of that being code, and utilizing synthetic data refined through Qwen2.5-Coder to enhance both coding skills and overall capabilities. Furthermore, the model benefits from post-training techniques that leverage extensive, execution-guided reinforcement learning, which facilitates the generation of diverse test cases across 20,000 parallel environments, thereby excelling in multi-turn software engineering tasks such as SWE-Bench Verified without needing test-time scaling. In addition to the model itself, the open-source Qwen Code CLI, derived from Gemini Code, empowers users to deploy Qwen3-Coder in dynamic workflows with tailored prompts and function calling protocols, while also offering smooth integration with Node.js, OpenAI SDKs, and environment variables. This comprehensive ecosystem supports developers in optimizing their coding projects effectively and efficiently.

The Mixtral 8x7B model is an advanced sparse mixture of experts (SMoE) system that boasts open weights and is released under the Apache 2.0 license. This model demonstrates superior performance compared to Llama 2 70B across various benchmarks while achieving inference speeds that are six times faster. Recognized as the leading open-weight model with a flexible licensing framework, Mixtral also excels in terms of cost-efficiency and performance. Notably, it competes with and often surpasses GPT-3.5 in numerous established benchmarks, highlighting its significance in the field. Its combination of accessibility, speed, and effectiveness makes it a compelling choice for developers seeking high-performing AI solutions.

We offer fast and precise AI models optimized for deployment in production environments. Our inference API is designed for high availability, utilizing cutting-edge NVIDIA GPUs to ensure optimal performance. We have curated a selection of top open-source natural language processing (NLP) models from the community, making them readily available for your use. You have the flexibility to fine-tune your own models, including GPT-J, or upload your proprietary models for seamless deployment in production. From your user-friendly dashboard, you can easily upload or train/fine-tune AI models, allowing you to integrate them into production immediately without the hassle of managing deployment factors such as memory usage, availability, or scalability. Moreover, you can upload an unlimited number of models and deploy them as needed, ensuring that you can continuously innovate and adapt to your evolving requirements. This provides a robust framework for leveraging AI technologies in your projects.

Qwen2.5-1M, an open-source language model from the Qwen team, has been meticulously crafted to manage context lengths reaching as high as one million tokens. This version introduces two distinct model variants, namely Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, representing a significant advancement as it is the first instance of Qwen models being enhanced to accommodate such large context lengths. In addition to this, the team has released an inference framework that is based on vLLM and incorporates sparse attention mechanisms, which greatly enhance the processing speed for 1M-token inputs, achieving improvements between three to seven times. A detailed technical report accompanies this release, providing in-depth insights into the design choices and the results from various ablation studies. This transparency allows users to fully understand the capabilities and underlying technology of the models.

This repository showcases the research preview of LongLLaMA, an advanced large language model that can manage extensive contexts of up to 256,000 tokens or potentially more. LongLLaMA is developed on the OpenLLaMA framework and has been fine-tuned utilizing the Focused Transformer (FoT) technique. The underlying code for LongLLaMA is derived from Code Llama. We are releasing a smaller 3B base variant of the LongLLaMA model, which is not instruction-tuned, under an open license (Apache 2.0), along with inference code that accommodates longer contexts available on Hugging Face. This model's weights can seamlessly replace LLaMA in existing systems designed for shorter contexts, specifically those handling up to 2048 tokens. Furthermore, we include evaluation results along with comparisons to the original OpenLLaMA models, thereby providing a comprehensive overview of LongLLaMA's capabilities in the realm of long-context processing.

CodeQwen serves as the coding counterpart to Qwen, which is a series of large language models created by the Qwen team at Alibaba Cloud. Built on a transformer architecture that functions solely as a decoder, this model has undergone extensive pre-training using a vast dataset of code. It showcases robust code generation abilities and demonstrates impressive results across various benchmarking tests. With the capacity to comprehend and generate long contexts of up to 64,000 tokens, CodeQwen accommodates 92 programming languages and excels in tasks such as text-to-SQL queries and debugging. Engaging with CodeQwen is straightforward—you can initiate a conversation with just a few lines of code utilizing transformers. The foundation of this interaction relies on constructing the tokenizer and model using pre-existing methods, employing the generate function to facilitate dialogue guided by the chat template provided by the tokenizer. In alignment with our established practices, we implement the ChatML template tailored for chat models. This model adeptly completes code snippets based on the prompts it receives, delivering responses without the need for any further formatting adjustments, thereby enhancing the user experience. The seamless integration of these elements underscores the efficiency and versatility of CodeQwen in handling diverse coding tasks.

Falcon-40B is a causal decoder-only model consisting of 40 billion parameters, developed by TII and trained on 1 trillion tokens from RefinedWeb, supplemented with carefully selected datasets. It is distributed under the Apache 2.0 license. Why should you consider using Falcon-40B? This model stands out as the leading open-source option available, surpassing competitors like LLaMA, StableLM, RedPajama, and MPT, as evidenced by its ranking on the OpenLLM Leaderboard. Its design is specifically tailored for efficient inference, incorporating features such as FlashAttention and multiquery capabilities. Moreover, it is offered under a flexible Apache 2.0 license, permitting commercial applications without incurring royalties or facing restrictions. It's important to note that this is a raw, pretrained model and is generally recommended to be fine-tuned for optimal performance in most applications. If you need a version that is more adept at handling general instructions in a conversational format, you might want to explore Falcon-40B-Instruct as a potential alternative.

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.

Falcon-7B is a causal decoder-only model comprising 7 billion parameters, developed by TII and trained on an extensive dataset of 1,500 billion tokens from RefinedWeb, supplemented with specially selected corpora, and it is licensed under Apache 2.0. What are the advantages of utilizing Falcon-7B? This model surpasses similar open-source alternatives, such as MPT-7B, StableLM, and RedPajama, due to its training on a remarkably large dataset of 1,500 billion tokens from RefinedWeb, which is further enhanced with carefully curated content, as evidenced by its standing on the OpenLLM Leaderboard. Additionally, it boasts an architecture that is finely tuned for efficient inference, incorporating technologies like FlashAttention and multiquery mechanisms. Moreover, the permissive nature of the Apache 2.0 license means users can engage in commercial applications without incurring royalties or facing significant limitations. This combination of performance and flexibility makes Falcon-7B a strong choice for developers seeking advanced modeling capabilities.

Mistral AI has unveiled two cutting-edge models specifically designed for on-device computing and edge use cases, collectively referred to as "les Ministraux": Ministral 3B and Ministral 8B. These innovative models stand out due to their capabilities in knowledge retention, commonsense reasoning, function-calling, and overall efficiency, all while remaining within the sub-10B parameter range. They boast support for a context length of up to 128k, making them suitable for a diverse range of applications such as on-device translation, offline smart assistants, local analytics, and autonomous robotics. Notably, Ministral 8B incorporates an interleaved sliding-window attention mechanism, which enhances both the speed and memory efficiency of inference processes. Both models are adept at serving as intermediaries in complex multi-step workflows, skillfully managing functions like input parsing, task routing, and API interactions based on user intent, all while minimizing latency and operational costs. Benchmark results reveal that les Ministraux consistently exceed the performance of similar models across a variety of tasks, solidifying their position in the market. As of October 16, 2024, these models are now available for developers and businesses, with Ministral 8B being offered at a competitive rate of $0.1 for every million tokens utilized. This pricing structure enhances accessibility for users looking to integrate advanced AI capabilities into their solutions.

Recent breakthroughs in natural language processing, comprehension, and generation have been greatly influenced by the development of large language models. This research presents a system that employs Ascend 910 AI processors and the MindSpore framework to train a language model exceeding one trillion parameters, specifically 1.085 trillion, referred to as PanGu-{\Sigma}. This model enhances the groundwork established by PanGu-{\alpha} by converting the conventional dense Transformer model into a sparse format through a method known as Random Routed Experts (RRE). Utilizing a substantial dataset of 329 billion tokens, the model was effectively trained using a strategy called Expert Computation and Storage Separation (ECSS), which resulted in a remarkable 6.3-fold improvement in training throughput through the use of heterogeneous computing. Through various experiments, it was found that PanGu-{\Sigma} achieves a new benchmark in zero-shot learning across multiple downstream tasks in Chinese NLP, showcasing its potential in advancing the field. This advancement signifies a major leap forward in the capabilities of language models, illustrating the impact of innovative training techniques and architectural modifications.

In honor of Cleopatra, whose magnificent fate concluded amidst the tragic incident involving a snake, we are excited to introduce Codestral Mamba, a Mamba2 language model specifically designed for code generation and released under an Apache 2.0 license. Codestral Mamba represents a significant advancement in our ongoing initiative to explore and develop innovative architectures. It is freely accessible for use, modification, and distribution, and we aspire for it to unlock new avenues in architectural research. The Mamba models are distinguished by their linear time inference capabilities and their theoretical potential to handle sequences of infinite length. This feature enables users to interact with the model effectively, providing rapid responses regardless of input size. Such efficiency is particularly advantageous for enhancing code productivity; therefore, we have equipped this model with sophisticated coding and reasoning skills, allowing it to perform competitively with state-of-the-art transformer-based models. As we continue to innovate, we believe Codestral Mamba will inspire further advancements in the coding community.

NVIDIA NeMo Megatron serves as a comprehensive framework designed for the training and deployment of large language models (LLMs) that can range from billions to trillions of parameters. As a integral component of the NVIDIA AI platform, it provides a streamlined, efficient, and cost-effective solution in a containerized format for constructing and deploying LLMs. Tailored for enterprise application development, the framework leverages cutting-edge technologies stemming from NVIDIA research and offers a complete workflow that automates distributed data processing, facilitates the training of large-scale custom models like GPT-3, T5, and multilingual T5 (mT5), and supports model deployment for large-scale inference. The process of utilizing LLMs becomes straightforward with the availability of validated recipes and predefined configurations that streamline both training and inference. Additionally, the hyperparameter optimization tool simplifies the customization of models by automatically exploring the optimal hyperparameter configurations, enhancing performance for training and inference across various distributed GPU cluster setups. This approach not only saves time but also ensures that users can achieve superior results with minimal effort.

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.

Reka Flash 3 is a cutting-edge multimodal AI model with 21 billion parameters, crafted by Reka AI to perform exceptionally well in tasks such as general conversation, coding, following instructions, and executing functions. This model adeptly handles and analyzes a myriad of inputs, including text, images, video, and audio, providing a versatile and compact solution for a wide range of applications. Built from the ground up, Reka Flash 3 was trained on a rich array of datasets, encompassing both publicly available and synthetic information, and it underwent a meticulous instruction tuning process with high-quality selected data to fine-tune its capabilities. The final phase of its training involved employing reinforcement learning techniques, specifically using the REINFORCE Leave One-Out (RLOO) method, which combined both model-based and rule-based rewards to significantly improve its reasoning skills. With an impressive context length of 32,000 tokens, Reka Flash 3 competes effectively with proprietary models like OpenAI's o1-mini, making it an excellent choice for applications requiring low latency or on-device processing. The model operates at full precision with a memory requirement of 39GB (fp16), although it can be efficiently reduced to just 11GB through the use of 4-bit quantization, demonstrating its adaptability for various deployment scenarios. Overall, Reka Flash 3 represents a significant advancement in multimodal AI technology, capable of meeting diverse user needs across multiple platforms.

Our advanced multimodal assistant is meticulously crafted with a focus on privacy, security, and operational efficiency. Yasa is trained to interpret various forms of content, including text, images, videos, and tabular data, with plans to expand to additional modalities in the future. It can assist you in brainstorming for creative projects, answering fundamental questions, or extracting valuable insights from your internal datasets. With just a few straightforward commands, you can generate, train, compress, or deploy it on your own servers. Our proprietary algorithms enable you to customize the model according to your specific data and requirements. We utilize innovative techniques that encompass retrieval, fine-tuning, self-supervised instruction tuning, and reinforcement learning to optimize our model based on your unique datasets, ensuring that it meets your operational needs effectively. In doing so, we aim to enhance user experience and deliver tailored solutions that drive productivity and innovation.

Xgen-small is a compact language model crafted by Salesforce AI Research that is tailored for enterprise use, offering efficient long-context capabilities at a manageable cost. It employs a combination of focused data curation, scalable pre-training, length extension, instruction fine-tuning, and reinforcement learning to address the intricate and high-volume inference needs of contemporary businesses. In contrast to conventional large models, Xgen-small excels in processing extensive contexts, allowing it to effectively synthesize insights from various sources such as internal documents, code bases, research articles, and real-time data feeds. With parameter sizes of 4B and 9B, it strikes a careful balance between cost efficiency, privacy protections, and comprehensive long-context comprehension, positioning itself as a reliable and sustainable option for large-scale Enterprise AI implementation. This innovative approach not only enhances operational efficiency but also empowers organizations to leverage AI effectively in their strategic initiatives.

The Holo2 model family from H Company offers a blend of affordability and high performance in vision-language models specifically designed for computer-based agents that can navigate, localize user interface elements, and function across web, desktop, and mobile platforms. This new series, which is available in sizes of 4 billion, 8 billion, and 30 billion parameters, builds upon the foundations laid by the earlier Holo1 and Holo1.5 models, ensuring strong grounding in user interfaces while making substantial improvements to navigation abilities. Utilizing a mixture-of-experts (MoE) architecture, the Holo2 models activate only the necessary parameters to maximize operational efficiency. These models have been trained on carefully curated datasets focused on localization and agent functionality, allowing them to seamlessly replace their predecessors. They provide support for effortless inference in environments compatible with Qwen3-VL models and can be easily incorporated into agentic workflows such as Surfer 2. In benchmark evaluations, the Holo2-30B-A3B model demonstrated impressive results, achieving 66.1% accuracy on the ScreenSpot-Pro test and 76.1% on the OSWorld-G benchmark, thereby establishing itself as the leader in the UI localization sector. Additionally, the advancements in the Holo2 models make them a compelling choice for developers looking to enhance the efficiency and performance of their applications.

Grok 4.1 Fast represents xAI’s leap forward in building highly capable agents that rely heavily on tool calling, long-context reasoning, and real-time information retrieval. It supports a robust 2-million-token window, enabling long-form planning, deep research, and multi-step workflows without degradation. Through extensive RL training and exposure to diverse tool ecosystems, the model performs exceptionally well on demanding benchmarks like τ²-bench Telecom. When paired with the Agent Tools API, it can autonomously browse the web, search X posts, execute Python code, and retrieve documents, eliminating the need for developers to manage external infrastructure. It is engineered to maintain intelligence across multi-turn conversations, making it ideal for enterprise tasks that require continuous context. Its benchmark accuracy on tool-calling and function-calling tasks clearly surpasses competing models in speed, cost, and reliability. Developers can leverage these strengths to build agents that automate customer support, perform real-time analysis, and execute complex domain-specific tasks. With its performance, low pricing, and availability on platforms like OpenRouter, Grok 4.1 Fast stands out as a production-ready solution for next-generation AI systems.

An advanced End-to-End MLLM is designed to accept various forms of references and effectively ground responses. The Ferret Model utilizes a combination of Hybrid Region Representation and a Spatial-aware Visual Sampler, which allows for detailed and flexible referring and grounding capabilities within the MLLM framework. The GRIT Dataset, comprising approximately 1.1 million entries, serves as a large-scale and hierarchical dataset specifically crafted for robust instruction tuning in the ground-and-refer category. Additionally, the Ferret-Bench is a comprehensive multimodal evaluation benchmark that simultaneously assesses referring, grounding, semantics, knowledge, and reasoning, ensuring a well-rounded evaluation of the model's capabilities. This intricate setup aims to enhance the interaction between language and visual data, paving the way for more intuitive AI systems.

Mistral Small 3.1 represents a cutting-edge, multimodal, and multilingual AI model that has been released under the Apache 2.0 license. This upgraded version builds on Mistral Small 3, featuring enhanced text capabilities and superior multimodal comprehension, while also accommodating an extended context window of up to 128,000 tokens. It demonstrates superior performance compared to similar models such as Gemma 3 and GPT-4o Mini, achieving impressive inference speeds of 150 tokens per second. Tailored for adaptability, Mistral Small 3.1 shines in a variety of applications, including instruction following, conversational support, image analysis, and function execution, making it ideal for both business and consumer AI needs. The model's streamlined architecture enables it to operate efficiently on hardware such as a single RTX 4090 or a Mac equipped with 32GB of RAM, thus supporting on-device implementations. Users can download it from Hugging Face and access it through Mistral AI's developer playground, while it is also integrated into platforms like Google Cloud Vertex AI, with additional accessibility on NVIDIA NIM and more. This flexibility ensures that developers can leverage its capabilities across diverse environments and applications.