Alternatives to Tülu 3

NVIDIA Nemotron 3.5 Super VL is a multimodal foundation model developed for AI agents that require reasoning, visual understanding, coding, and multistep task execution. Its 120-billion-parameter mixture-of-experts architecture activates approximately 12 billion parameters per token, enabling efficient inference while maintaining substantial model capacity. The model processes text, images, documents, and video within a unified system that combines multimodal interpretation with instruction following and tool use. Its document intelligence capabilities include extracting information from tables, analyzing charts, interpreting page layouts, and processing multipage documents. Video understanding features support visual information retrieval, content review, incident analysis, and reasoning across recorded sequences. The model also supports software engineering tasks such as code analysis, error identification, debugging, testing, and recommending fixes. Developers can customize Nemotron 3.5 Super VL using NVIDIA NeMo for supervised fine-tuning, low-rank adaptation, and reinforcement learning with domain-specific data. Its inference optimizations include speculative decoding, context reuse, multi-token prediction, and NVFP4 quantization to improve throughput and reduce computational requirements. The model is available with open weights and supports deployment across compatible local hardware, enterprise data centers, and cloud infrastructure.

Muse Spark Open is an open-weight foundation model developed by Meta to support multimodal reasoning, autonomous agents, and complex software development tasks. The model features 816 billion total parameters with approximately 42 billion active parameters and a context window supporting up to one million tokens. Its architecture is designed to process multimodal information while maintaining context across lengthy workflows, extensive codebases, and multistep assignments. Muse Spark Open provides capabilities for software engineering, code generation, debugging, application development, and automated testing. Its agentic functionality supports external tool interactions, task planning, computer-based workflows, and execution of complex objectives over extended sessions. The model can run locally using two Apple M5 Ultra Mac Studio systems with the MLX framework, enabling organizations to process information without relying on hosted inference services. Developers can deploy the model through SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and other supported inference engines. Customization is supported through PyTorch and TorchTitan, allowing organizations to adapt the model to specialized applications and workflows. Muse Spark Open is distributed under the Muse Model License and is available through Hugging Face, major cloud model repositories, and hosted API providers including Together AI, Fireworks AI, and OpenRouter.

GLM-5.3 is Z.ai’s advanced coding and agentic reasoning model built through scaled post-training on top of the GLM-5.2 base model. The release focuses on frontier coding, long-horizon software engineering, agent tasks, cyber evaluation, and reinforcement learning at scale. GLM-5.3 improves significantly over GLM-5.2 on complex coding benchmarks, real-world engineering environments, Terminal Bench 3.0, DeepSWE, Agents’ Last Exam, and Z.ai’s internal Code Bench. The model is trained on environments that resemble real professional work, including tasks involving codebases, infrastructure, documentation, compute clusters, experiments, bottleneck diagnosis, implementation, testing, and measurable optimization. Z.ai’s post-training stack includes IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for large-scale asynchronous training. GLM-5.3 supports three thinking effort levels, including low, high, and max, with max recommended for coding tasks. The model also demonstrates emergent cyber capabilities across vulnerability discovery and exploitation benchmarks, prompting continued safety evaluation and hardening before weights are released. GLM-5.3 can be used through the GLM Coding Plan, ZCode, Claude Code, OpenCode, and other coding agent workflows. By combining stronger coding performance, long-horizon task execution, post-training scale, cyber evaluation, reasoning effort controls, and coding-agent integrations, GLM-5.3 supports advanced developer and research workflows.

MiMo-V2.6-Pro is an open-source, natively omnimodal AI model from Xiaomi MiMo designed for coding, agentic workflows, visual creation, research, and computer-based tasks. It is the highest-capability model in the MiMo-V2.6 series and was trained using a large-scale reinforcement learning process focused on verifiable, complex tasks. The model can handle software engineering, automation, tool use, visual reasoning, and long-running agent workflows across multiple environments. Its multimodal capabilities extend beyond traditional coding into 3D scene generation, Blender modeling, embodied simulation, frontend development, presentation design, video creation, and music composition. MiMo-V2.6-Pro can take text, images, video, and other visual inputs and coordinate multi-agent workflows to produce and iteratively refine complex outputs. Research applications demonstrated by Xiaomi include materials discovery, literature analysis, computational simulation, hypothesis generation, and formalizing mathematical proofs in Lean. Xiaomi has released the model alongside its technical report, reinforcement learning environments, and RL code so researchers can study and reproduce the training approach. MiMo-V2.6-Pro is also available in an UltraSpeed configuration that provides substantially faster output for latency-sensitive workflows. Users can access the model through MiMo Desktop, AI Studio, MiMo Code, the MiMo API Platform, OpenRouter, and Hugging Face.

Inkling is Thinking Machines’ open-weights foundation model built for customization, multimodal reasoning, and agentic AI workflows. The model uses a Mixture-of-Experts architecture with 975 billion total parameters and 41 billion active parameters, making it large in capacity while activating only a subset of experts per token. Inkling supports up to a 1 million token context window and was pretrained on 45 trillion tokens spanning text, images, audio, and video. It is designed as a broad generalist model with strengths across coding, reasoning, instruction following, factuality, tool use, vision, audio understanding, forecasting, and safety. Developers can tune its thinking effort to trade off latency, cost, and performance, which is useful for production systems that need efficient reasoning at scale. Inkling can be fine-tuned on Tinker, tested in the Inkling Playground, and deployed through partners such as TogetherAI, Fireworks, Modal, Databricks, Baseten, vLLM, SGLang, llama.cpp, and Hugging Face transformers. The model can generate applications, operate tools, create styled artifacts, reason over visual and audio inputs, and support long refinement loops for collaborative work. Thinking Machines also previewed Inkling-Small, a lighter Mixture-of-Experts model with 276 billion total parameters and 12 billion active parameters for lower-cost and lower-latency workloads. By combining open weights, multimodal training, agentic capabilities, efficient reasoning, and fine-tuning support, Inkling gives builders a flexible AI foundation for specialized products and workflows.

Qwen3.8-Max is a frontier AI model from the Qwen family designed for advanced coding, coworking, research, multimodal reasoning, and long-horizon autonomous tasks. It is described as Qwen’s most capable model to date and the first Qwen-Max-class model with open weights announced for release. The model scales to 2.4 trillion parameters with 95 billion active parameters and is accessible through QwenCloud. Qwen3.8-Max is built to answer difficult questions and complete complex deliverables from start to finish. Its coding capabilities include autonomous project creation, self-testing, issue dispatch, CI validation, pull request workflows, and long-running feedback loops. The model is also designed for real-world work across legal review, UI/UX design, restaurant operations, structural engineering, rehabilitation visualization, sports analytics, and quantitative research. Its multimodal capabilities support images, documents, videos, interface reconstruction, visual production, application recreation, and visual feedback loops. QwenCloud supports industry-standard API protocols, including OpenAI-compatible chat completions and responses APIs as well as an Anthropic-compatible interface. By combining large-scale reasoning, agentic coding, multimodal intelligence, API access, long-context workflows, and open-weight availability, Qwen3.8-Max gives teams a powerful foundation for building advanced AI systems.

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.

Molmo represents a cutting-edge family of multimodal AI models crafted by the Allen Institute for AI (Ai2). These innovative models are specifically engineered to connect the divide between open-source and proprietary systems, ensuring they perform competitively across numerous academic benchmarks and assessments by humans. In contrast to many existing multimodal systems that depend on synthetic data sourced from proprietary frameworks, Molmo is exclusively trained on openly available data, which promotes transparency and reproducibility in AI research. A significant breakthrough in the development of Molmo is the incorporation of PixMo, a unique dataset filled with intricately detailed image captions gathered from human annotators who utilized speech-based descriptions, along with 2D pointing data that empowers the models to respond to inquiries with both natural language and non-verbal signals. This capability allows Molmo to engage with its surroundings in a more sophisticated manner, such as by pointing to specific objects within images, thereby broadening its potential applications in diverse fields, including robotics, augmented reality, and interactive user interfaces. Furthermore, the advancements made by Molmo set a new standard for future multimodal AI research and application development.

Mistral 7B is a language model with 7.3 billion parameters that demonstrates superior performance compared to larger models such as Llama 2 13B on a variety of benchmarks. It utilizes innovative techniques like Grouped-Query Attention (GQA) for improved inference speed and Sliding Window Attention (SWA) to manage lengthy sequences efficiently. Released under the Apache 2.0 license, Mistral 7B is readily available for deployment on different platforms, including both local setups and prominent cloud services. Furthermore, a specialized variant known as Mistral 7B Instruct has shown remarkable capabilities in following instructions, outperforming competitors like Llama 2 13B Chat in specific tasks. This versatility makes Mistral 7B an attractive option for developers and researchers alike.

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 NVIDIA Llama Nemotron family comprises a series of sophisticated language models that are fine-tuned for complex reasoning and a wide array of agentic AI applications. These models shine in areas such as advanced scientific reasoning, complex mathematics, coding, following instructions, and executing tool calls. They are designed for versatility, making them suitable for deployment on various platforms, including data centers and personal computers, and feature the ability to switch reasoning capabilities on or off, which helps to lower inference costs during less demanding tasks. The Llama Nemotron series consists of models specifically designed to meet different deployment requirements. Leveraging the foundation of Llama models and enhanced through NVIDIA's post-training techniques, these models boast a notable accuracy improvement of up to 20% compared to their base counterparts while also achieving inference speeds that can be up to five times faster than other leading open reasoning models. This remarkable efficiency allows for the management of more intricate reasoning challenges, boosts decision-making processes, and significantly lowers operational expenses for businesses. Consequently, the Llama Nemotron models represent a significant advancement in the field of AI, particularly for organizations seeking to integrate cutting-edge reasoning capabilities into their systems.

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.

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.

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.

The latest iteration of the open-source AI model, which can be fine-tuned and deployed in various environments, is now offered in multiple versions, including 1B, 3B, 11B, and 90B, alongside the option to continue utilizing Llama 3.1. Llama 3.2 comprises a series of large language models (LLMs) that come pretrained and fine-tuned in 1B and 3B configurations for multilingual text only, while the 11B and 90B models accommodate both text and image inputs, producing text outputs. With this new release, you can create highly effective and efficient applications tailored to your needs. For on-device applications, such as summarizing phone discussions or accessing calendar tools, the 1B or 3B models are ideal choices. Meanwhile, the 11B or 90B models excel in image-related tasks, enabling you to transform existing images or extract additional information from images of your environment. Overall, this diverse range of models allows developers to explore innovative use cases across various domains.

Mistral Large 3 pushes open-source AI into frontier territory with a massive sparse MoE architecture that activates 41B parameters per token while maintaining a highly efficient 675B total parameter design. It sets a new performance standard by combining long-context reasoning, multilingual fluency across 40+ languages, and robust multimodal comprehension within a single unified model. Trained end-to-end on thousands of NVIDIA H200 GPUs, it reaches parity with top closed-source instruction models while remaining fully accessible under the Apache 2.0 license. Developers benefit from optimized deployments through partnerships with NVIDIA, Red Hat, and vLLM, enabling smooth inference on A100, H100, and Blackwell-class systems. The model ships in both base and instruct variants, with a reasoning-enhanced version on the way for even deeper analytical capabilities. Beyond general intelligence, Mistral Large 3 is engineered for enterprise customization, allowing organizations to refine the model on internal datasets or domain-specific tasks. Its efficient token generation and powerful multimodal stack make it ideal for coding, document analysis, knowledge workflows, agentic systems, and multilingual communications. With Mistral Large 3, organizations can finally deploy frontier-class intelligence with full transparency, flexibility, and control.

Smaug-72B is a formidable open-source large language model (LLM) distinguished by several prominent features: Exceptional Performance: It currently ranks first on the Hugging Face Open LLM leaderboard, outperforming models such as GPT-3.5 in multiple evaluations, demonstrating its ability to comprehend, react to, and generate text that closely resembles human writing. Open Source Availability: In contrast to many high-end LLMs, Smaug-72B is accessible to everyone for use and modification, which encourages cooperation and innovation within the AI ecosystem. Emphasis on Reasoning and Mathematics: This model excels particularly in reasoning and mathematical challenges, a capability attributed to specialized fine-tuning methods developed by its creators, Abacus AI. Derived from Qwen-72B: It is essentially a refined version of another robust LLM, Qwen-72B, which was launched by Alibaba, thereby enhancing its overall performance. In summary, Smaug-72B marks a notable advancement in the realm of open-source artificial intelligence, making it a valuable resource for developers and researchers alike. Its unique strengths not only elevate its status but also contribute to the ongoing evolution of AI technology.

Phi-4-reasoning is an advanced transformer model featuring 14 billion parameters, specifically tailored for tackling intricate reasoning challenges, including mathematics, programming, algorithm development, and strategic planning. Through a meticulous process of supervised fine-tuning on select "teachable" prompts and reasoning examples created using o3-mini, it excels at generating thorough reasoning sequences that optimize computational resources during inference. By integrating outcome-driven reinforcement learning, Phi-4-reasoning is capable of producing extended reasoning paths. Its performance notably surpasses that of significantly larger open-weight models like DeepSeek-R1-Distill-Llama-70B and nears the capabilities of the comprehensive DeepSeek-R1 model across various reasoning applications. Designed for use in settings with limited computing power or high latency, Phi-4-reasoning is fine-tuned with synthetic data provided by DeepSeek-R1, ensuring it delivers precise and methodical problem-solving. This model's ability to handle complex tasks with efficiency makes it a valuable tool in numerous computational contexts.

Liquid AI's LFM2.5 represents an advanced iteration of on-device AI foundation models, engineered to provide high-efficiency and performance for AI inference on edge devices like smartphones, laptops, vehicles, IoT systems, and embedded hardware without the need for cloud computing resources. This new version builds upon the earlier LFM2 framework by greatly enhancing the scale of pretraining and the stages of reinforcement learning, resulting in a suite of hybrid models that boast around 1.2 billion parameters while effectively balancing instruction adherence, reasoning skills, and multimodal functionalities for practical applications. The LFM2.5 series comprises various models including Base (for fine-tuning and personalization), Instruct (designed for general-purpose instruction), Japanese-optimized, Vision-Language, and Audio-Language variants, all meticulously crafted for rapid on-device inference even with stringent memory limitations. These models are also made available as open-weight options, facilitating deployment through platforms such as llama.cpp, MLX, vLLM, and ONNX, thus ensuring versatility for developers. With these enhancements, LFM2.5 positions itself as a robust solution for diverse AI-driven tasks in real-world environments.

EXAONE Deep represents a collection of advanced language models that are enhanced for reasoning, created by LG AI Research, and come in sizes of 2.4 billion, 7.8 billion, and 32 billion parameters. These models excel in a variety of reasoning challenges, particularly in areas such as mathematics and coding assessments. Significantly, the EXAONE Deep 2.4B model outshines other models of its size, while the 7.8B variant outperforms both open-weight models of similar dimensions and the proprietary reasoning model known as OpenAI o1-mini. Furthermore, the EXAONE Deep 32B model competes effectively with top-tier open-weight models in the field. The accompanying repository offers extensive documentation that includes performance assessments, quick-start guides for leveraging EXAONE Deep models with the Transformers library, detailed explanations of quantized EXAONE Deep weights formatted in AWQ and GGUF, as well as guidance on how to run these models locally through platforms like llama.cpp and Ollama. Additionally, this resource serves to enhance user understanding and accessibility to the capabilities of EXAONE Deep models.

Laguna XS.2 represents Poolside’s innovative open-weight coding model, distinguished as the lightest and quickest member of the Laguna series. This model features a total of 33 billion parameters in a Mixture of Experts setup, with 3 billion parameters activated, and has been meticulously trained in-house using 30 trillion tokens. As the latest generation model accessible to the public, it embodies a second-generation architecture and marks Poolside’s inaugural open-weight offering, drawing from insights gained during the training of Laguna M.1 with synthetic data and reinforcement learning techniques. Specifically designed to enhance agentic coding workflows, Laguna XS.2 excels in coding, acting, and rapidly iterating, particularly within Poolside’s coding agent environment. This model is particularly advantageous for developers and teams seeking a lightweight, efficient coding solution rather than a more cumbersome frontier system. Released under the permissive Apache 2.0 license, it empowers the community to assess, fine-tune, quantize, and build upon its weights, fostering a collaborative development atmosphere. In essence, Laguna XS.2 not only provides a robust platform for agentic coding but also encourages innovation and experimentation among its users.

Code Llama is an advanced language model designed to generate code through text prompts, distinguishing itself as a leading tool among publicly accessible models for coding tasks. This innovative model not only streamlines workflows for existing developers but also aids beginners in overcoming challenges associated with learning to code. Its versatility positions Code Llama as both a valuable productivity enhancer and an educational resource, assisting programmers in creating more robust and well-documented software solutions. Additionally, users can generate both code and natural language explanations by providing either type of prompt, making it an adaptable tool for various programming needs. Available for free for both research and commercial applications, Code Llama is built upon Llama 2 architecture and comes in three distinct versions: the foundational Code Llama model, Code Llama - Python which is tailored specifically for Python programming, and Code Llama - Instruct, optimized for comprehending and executing natural language directives effectively.

Fastino operates as an applied AI platform that specializes in open-weight language models along with the Fastino Fine-Tuning Agent. This innovative agent allows users to articulate a task using simple language, subsequently determining the appropriate architecture, generating the necessary training data, conducting training and evaluation, and ultimately delivering a task-specific model that is ready for deployment. Users can initiate and revisit fine-tuning projects through a single interface, ensuring that models are developed according to their specifications and can be deployed within their own environments. The models produced by Fastino are tailored for production-grade efficiency, typically achieving response times of less than 50 milliseconds, while maintaining user ownership and privacy of the model weights. Notably, models can transition from a basic task description to a fully trained output in mere hours, enabling teams to expedite their specialized deployment processes significantly. Additionally, Fastino offers a selection of open-source and open-weight models specifically designed for various specialized AI applications, further enhancing accessibility and versatility for users.

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.

Qwen3.8-27B is a 27B-class open-weights model associated with Alibaba’s Qwen3.8 model family. Alibaba’s Qwen3.8 release positioned the broader family as a top-tier large language model system optimized for coding and professional cowork scenarios. Reports indicate that Qwen3.8-27B was planned to be released as open weights alongside Qwen3.8-Max, giving developers and researchers a more accessible option than the full Max-scale model. The model is designed for users who want strong AI capability in a smaller, more deployable package. Qwen3.8-27B can support workflows such as coding assistance, AI agents, research tasks, document analysis, data work, and self-hosted experimentation. The larger Qwen3.8-Max release is described as targeting coding, research, professional work, and multimodal tasks, and Qwen3.8-27B appears to serve builders who need a more practical model size for local or private infrastructure. QwenCloud documentation confirms that the Qwen3.8 generation includes modern capabilities such as thinking, function calling, built-in tools, and structured output for the Max model. Community discussion and third-party coverage also highlight interest in running Qwen3.8-27B through GGUF and local inference workflows. By combining open-weight accessibility, a 27B-class footprint, Qwen3.8-era capability, and developer-focused use cases, Qwen3.8-27B gives teams a practical model for coding and agentic experimentation.

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.

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 are excited to unveil Codestral, our inaugural code generation model. This open-weight generative AI system is specifically crafted for tasks related to code generation, enabling developers to seamlessly write and engage with code via a unified instruction and completion API endpoint. As it becomes proficient in both programming languages and English, Codestral is poised to facilitate the creation of sophisticated AI applications tailored for software developers. With a training foundation that encompasses a wide array of over 80 programming languages—ranging from widely-used options like Python, Java, C, C++, JavaScript, and Bash to more niche languages such as Swift and Fortran—Codestral ensures a versatile support system for developers tackling various coding challenges and projects. Its extensive language capabilities empower developers to confidently navigate different coding environments, making Codestral an invaluable asset in the programming landscape.

OpenPipe offers an efficient platform for developers to fine-tune their models. It allows you to keep your datasets, models, and evaluations organized in a single location. You can train new models effortlessly with just a click. The system automatically logs all LLM requests and responses for easy reference. You can create datasets from the data you've captured, and even train multiple base models using the same dataset simultaneously. Our managed endpoints are designed to handle millions of requests seamlessly. Additionally, you can write evaluations and compare the outputs of different models side by side for better insights. A few simple lines of code can get you started; just swap out your Python or Javascript OpenAI SDK with an OpenPipe API key. Enhance the searchability of your data by using custom tags. Notably, smaller specialized models are significantly cheaper to operate compared to large multipurpose LLMs. Transitioning from prompts to models can be achieved in minutes instead of weeks. Our fine-tuned Mistral and Llama 2 models routinely exceed the performance of GPT-4-1106-Turbo, while also being more cost-effective. With a commitment to open-source, we provide access to many of the base models we utilize. When you fine-tune Mistral and Llama 2, you maintain ownership of your weights and can download them whenever needed. Embrace the future of model training and deployment with OpenPipe's comprehensive tools and features.

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.

Qwen3.6 is an advanced AI model from Alibaba that builds on previous Qwen releases with a focus on real-world utility and performance. It is designed as a multimodal large language model capable of understanding and generating text while also processing visual and structured data. The model is optimized for coding tasks, enabling developers to handle complex, repository-level programming workflows. Qwen3.6 uses a mixture-of-experts (MoE) architecture, which activates only a portion of its parameters during inference to improve efficiency. This design allows it to deliver strong performance while reducing computational costs. It is available in both proprietary and open-weight versions, giving developers flexibility in deployment. The model supports integration into enterprise systems and cloud platforms, particularly within Alibaba’s ecosystem. Qwen3.6 also introduces stronger agentic capabilities, allowing it to perform multi-step reasoning and more autonomous task execution. It is designed to handle complex workflows, including engineering, analysis, and decision-making tasks. The model emphasizes stability and responsiveness based on developer feedback. Overall, Qwen3.6 provides a scalable and efficient AI solution for coding, automation, and multimodal applications.

Qwen3.8-Flash-Next represents an open-weight multimodal Mixture-of-Experts architecture and serves as an initial glimpse into the design intended for Qwen4. This model strategically enhances attention mechanisms, residual pathways, embeddings, and optimization techniques to boost its capabilities, improve computational efficiency, expand model capacity, and ensure training stability. Its innovative hybrid architecture merges Gated DeltaNet, which adeptly compresses past information, with Qwen Sparse Attention, enabling the selection of significant context at a micro-block level to lessen both attention and indexing costs associated with lengthy sequences. The Gated Residual feature broadens the residual pathway into four streams, dynamically managing the flow of information across different layers. Additionally, the N-gram Embedding integrates large-scale local-pattern memory with minimal added computation per token, and it can be transferred to host memory for further efficiency. The model is structured around a 125B-parameter main network supplemented by 51B parameters dedicated to N-gram embeddings, activating only 6B parameters for each token processed. This sophisticated framework highlights the ongoing advancements in machine learning architectures, setting a promising stage for future developments.

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.

Instruction-following models like GPT-3.5 (text-DaVinci-003), ChatGPT, Claude, and Bing Chat have seen significant advancements in their capabilities, leading to a rise in their usage among individuals in both personal and professional contexts. Despite their growing popularity and integration into daily tasks, these models are not without their shortcomings, as they can sometimes disseminate inaccurate information, reinforce harmful stereotypes, and use inappropriate language. To effectively tackle these critical issues, it is essential for researchers and scholars to become actively involved in exploring these models further. However, conducting research on instruction-following models within academic settings has posed challenges due to the unavailability of models with comparable functionality to proprietary options like OpenAI’s text-DaVinci-003. In response to this gap, we are presenting our insights on an instruction-following language model named Alpaca, which has been fine-tuned from Meta’s LLaMA 7B model, aiming to contribute to the discourse and development in this field. This initiative represents a step towards enhancing the understanding and capabilities of instruction-following models in a more accessible manner for researchers.

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.

Z-Image is a family of open-source image generation foundation models created by Alibaba's Tongyi-MAI team, utilizing a Scalable Single-Stream Diffusion Transformer architecture to produce both photorealistic and imaginative images from textual descriptions with only 6 billion parameters, which enhances its efficiency compared to many larger models while maintaining competitive quality and responsiveness to instructions. This model family comprises several variants, including Z-Image-Turbo, a distilled version designed for rapid inference that achieves results with as few as eight function evaluations and sub-second generation times on compatible GPUs; Z-Image, the comprehensive foundation model tailored for high-fidelity creative outputs and fine-tuning processes; Z-Image-Omni-Base, a flexible base checkpoint aimed at fostering community-driven advancements; and Z-Image-Edit, specifically optimized for image-to-image editing tasks while demonstrating strong adherence to instructions. Each variant of Z-Image serves distinct purposes, catering to a wide range of user needs within the realm of image generation.

Devstral is a collaborative effort between Mistral AI and All Hands AI, resulting in an open-source large language model specifically tailored for software engineering. This model demonstrates remarkable proficiency in navigating intricate codebases, managing edits across numerous files, and addressing practical problems, achieving a notable score of 46.8% on the SWE-Bench Verified benchmark, which is superior to all other open-source models. Based on Mistral-Small-3.1, Devstral boasts an extensive context window supporting up to 128,000 tokens. It is designed for optimal performance on high-performance hardware setups, such as Macs equipped with 32GB of RAM or Nvidia RTX 4090 GPUs, and supports various inference frameworks including vLLM, Transformers, and Ollama. Released under the Apache 2.0 license, Devstral is freely accessible on platforms like Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio, allowing developers to integrate its capabilities into their projects seamlessly. This model not only enhances productivity for software engineers but also serves as a valuable resource for anyone working with code.

Smaug Flash encompasses a trio of open-weight models meticulously fine-tuned by Abacus.AI to address production agentic workloads, with each model strategically placed along the capability–efficiency spectrum. This model line is developed through a combination of human-curated real-world agentic data and synthetic examples grounded in challenging scenarios, which results in enhancements in agentic programming, real-world tool utilization, automation, long-context reasoning, and adherence to instructions. The flagship model, Smaug Flash, derived from DeepSeek V4 Flash 0731, serves as the primary solution for enterprise agents that require a harmonious blend of speed, efficiency, and dependable performance. Its specific tuning minimizes the potential for spins and confusion during extensive tool use while preserving the speed advantages of the base model. Additionally, Smaug Mini, built on Qwen3.8 27B, is designed for multimodal applications and smaller reasoning tasks, offering a more compact solution with improved real-world agentic capabilities for singular workflows. Together, these models cater to diverse operational needs across various applications, showcasing the versatility of the Smaug Flash family.

Tinker is an innovative training API tailored for researchers and developers, providing comprehensive control over model fine-tuning while simplifying the complexities of infrastructure management. It offers essential primitives that empower users to create bespoke training loops, supervision techniques, and reinforcement learning workflows. Currently, it facilitates LoRA fine-tuning on open-weight models from both the LLama and Qwen families, accommodating a range of model sizes from smaller variants to extensive mixture-of-experts configurations. Users can write Python scripts to manage data, loss functions, and algorithmic processes, while Tinker autonomously takes care of scheduling, resource distribution, distributed training, and recovery from failures. The platform allows users to download model weights at various checkpoints without the burden of managing the computational environment. Delivered as a managed service, Tinker executes training jobs on Thinking Machines’ proprietary GPU infrastructure, alleviating users from the challenges of cluster orchestration and enabling them to focus on building and optimizing their models. This seamless integration of capabilities makes Tinker a vital tool for advancing machine learning research and development.

GLM-5.1 represents the latest advancement in Z.ai’s GLM series, crafted as a cutting-edge, agent-focused AI model tailored for coding, reasoning, and managing long-term workflows. This iteration builds upon the framework of GLM-5, which employs a Mixture-of-Experts (MoE) architecture to achieve high performance without incurring excessive inference expenses, aligning with a larger initiative towards open-weight models that are accessible to developers. A significant emphasis of GLM-5.1 is on fostering agentic behavior, allowing it to plan, execute, and refine multi-step tasks instead of merely reacting to isolated prompts. Its capabilities are specifically engineered to manage intricate workflows, such as debugging code, exploring repositories, and performing sequential operations while maintaining context over time. In comparison to its predecessors, GLM-5.1 enhances reliability during lengthy interactions, ensuring coherence throughout extended sessions and minimizing failures in multi-step reasoning processes. Overall, this model signifies a leap forward in AI development, particularly in its ability to support complex task management seamlessly.

StableVicuna represents the inaugural large-scale open-source chatbot developed through reinforced learning from human feedback (RLHF). It is an advanced version of the Vicuna v0 13b model, which has undergone further instruction fine-tuning and RLHF training. To attain the impressive capabilities of StableVicuna, we use Vicuna as the foundational model and adhere to the established three-stage RLHF framework proposed by Steinnon et al. and Ouyang et al. Specifically, we perform additional training on the base Vicuna model with supervised fine-tuning (SFT), utilizing a blend of three distinct datasets. The first is the OpenAssistant Conversations Dataset (OASST1), which consists of 161,443 human-generated messages across 66,497 conversation trees in 35 languages. The second dataset is GPT4All Prompt Generations, encompassing 437,605 prompts paired with responses created by GPT-3.5 Turbo. Lastly, the Alpaca dataset features 52,000 instructions and demonstrations that were produced using OpenAI's text-davinci-003 model. This collective approach to training enhances the chatbot's ability to engage effectively in diverse conversational contexts.

Amazon Nova Micro is an advanced text-only AI model optimized for rapid language processing at a very low cost. With capabilities in reasoning, translation, and code completion, it offers over 200 tokens per second in response generation, making it suitable for fast-paced, real-time applications. Nova Micro supports fine-tuning with text inputs, and its efficiency in understanding and generating text makes it a cost-effective solution for AI-driven applications requiring high performance and quick outputs.

Push the limits of individual alignment, artificial consciousness, open-source software, and decentralization through experimentation that larger corporations and governments often shy away from. Hermes 3 features sophisticated long-term context retention, the ability to engage in multi-turn conversations, and intricate roleplaying and internal monologue capabilities, alongside improved functionality for agentic function-calling. The design of this model emphasizes precise adherence to system prompts and instruction sets in a flexible way. By fine-tuning Llama 3.1 across various scales, including 8B, 70B, and 405B, and utilizing a dataset largely composed of synthetically generated inputs, Hermes 3 showcases performance that rivals and even surpasses Llama 3.1, while also unlocking greater potential in reasoning and creative tasks. This series of instructive and tool-utilizing models exhibits exceptional reasoning and imaginative skills, paving the way for innovative applications. Ultimately, Hermes 3 represents a significant advancement in the landscape of AI development.

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.

Qwen3.8-2.4T-A95B stands out as the most extensive open model within the Qwen3.8 series, offering advanced Qwen-Max-class features in a publicly accessible format. Constructed upon the solid framework of Qwen3.5, this model significantly enhances performance in areas such as coding, professional tasks, research, and complex, prolonged agentic activities, emphasizing the reliability of executing intricate, multi-step workflows to completion. Utilizing a cutting-edge mixture-of-experts architecture, it boasts an impressive total of 2.4 trillion parameters, with 95 billion of those being activated, featuring 512 experts and engaging 10 routed along with one shared expert simultaneously. The model accommodates a native context length of 262,144 tokens, which can be extended to around 1.01 million tokens, thereby providing substantial flexibility for various applications. Furthermore, improvements in agent execution, such as enhanced autonomous planning and better responsiveness to environmental feedback, contribute to its efficiency, while its broader compatibility with widely used agent frameworks and development tools facilitates seamless integration into existing systems, making it a versatile choice for developers and researchers alike.