Top Tülu 3 Alternatives

NVIDIA Nemotron 3.5 Super VL is an open multimodal artificial intelligence model designed to support specialized autonomous agents, enterprise automation, and complex reasoning tasks. Introduced in October 2026, the model combines language processing, visual understanding, software engineering capabilities, and external tool interactions within a unified architecture. It uses a 120-billion-parameter mixture-of-experts design with approximately 12 billion active parameters per token to optimize inference efficiency. The model was distilled from NVIDIA Nemotron 3 Ultra and incorporates additional training focused on coding, multimodal reasoning, and agentic task execution. Its vision-language capabilities support analysis of images, charts, tables, diagrams, multipage documents, and video content. Document intelligence workflows can extract structured information, interpret visual evidence, connect information across pages, and generate summaries or reports. Software engineering features enable the model to analyze repositories, identify programming errors, test code, and recommend modifications across development workflows. Its agentic capabilities support web research, external tool usage, multistep problem-solving, cybersecurity vulnerability analysis, and enterprise process automation. NVIDIA NeMo provides customization tools for supervised fine-tuning, low-rank adaptation, and reinforcement learning using proprietary datasets, organizational policies, and specialized workflows. The model incorporates LiLiCorr speculative decoding, multi-token prediction, context reuse, and NVFP4 quantization to improve inference performance on compatible hardware. Nemotron 3.5 Super VL is distributed with open weights and supports local, cloud, and data center deployment through NVIDIA infrastructure and tools such as NVIDIA NIM.

Muse Spark Open is a large-scale open-weight artificial intelligence model developed by Meta for multimodal understanding, software engineering, autonomous agents, and long-horizon reasoning. Released in October 2026, the model provides downloadable weights that allow developers to deploy and customize it outside Meta's hosted AI services. It uses a mixture-of-experts architecture containing 816 billion total parameters, with approximately 42 billion parameters activated during inference. Its one-million-token context window supports processing extensive documents, large software repositories, and lengthy sequences of instructions and interactions. The model is designed for complex coding workflows involving source code analysis, debugging, feature implementation, application development, and software testing. Its multimodal capabilities enable AI agents to interpret visual information and combine that understanding with reasoning and automated computer interactions. Meta demonstrated the model performing local software development tasks and controlling DaVinci Resolve to analyze video footage and prepare edited content. Muse Spark Open supports on-device deployment using two Apple M5 Ultra Mac Studio computers running the MLX framework, with additional compatibility across AMD, Intel, and NVIDIA hardware. Supported inference frameworks include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth, with optimized GPU kernels for improved processing performance. Developers can customize the model using PyTorch and TorchTitan to adapt its behavior to specialized tasks, organizational requirements, and application workflows. Muse Spark Open is distributed under the Muse Model License with limited commercial terms and is available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models, and several hosted API providers.

Inkling is an open-weights multimodal AI model from Thinking Machines built to support customization, agentic workflows, coding, reasoning, vision, audio, and enterprise AI use cases. The model is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active parameters, 256 routed experts per MoE layer, and six routed experts active per token. It supports context windows up to 1 million tokens and was pretrained on 45 trillion tokens across text, images, audio, and video. Inkling is designed as a broad foundation model rather than a narrowly optimized benchmark model, giving it balanced capabilities across reasoning, coding, factuality, instruction following, vision, audio, tool use, and safety. Its controllable thinking effort lets developers adjust how much computation and generated reasoning the model uses, helping teams balance quality, latency, and cost for different production needs. The model can run agentic coding tasks, use tools, create web apps, generate polished multi-page artifacts, reason over long contexts, and work through iterative refinement loops. For multimodal tasks, Inkling can process images, answer questions about visual content, transcribe and reason over audio, follow spoken instructions, and combine visual reasoning with code-based tools such as Python. Thinking Machines trained Inkling for calibration, instruction following, factual reliability, refusal behavior, and safety across multiple modalities, including evaluations for dangerous capabilities and human-AI threat vectors. Inkling is available on Tinker for fine-tuning, with 64K and 256K context options, an Inkling Playground for testing, cookbook recipes, and support for multimodal post-training workflows. Its full weights are available on Hugging Face, and deployment support is available through APIs and infrastructure partners such as TogetherAI, Fireworks, Modal, Databricks, Baseten, SGLang, vLLM, llama.cpp, and transformers.

GLM-5.3 is Z.ai’s frontier coding model built to improve complex software engineering, long-horizon agent work, and advanced technical reasoning through scaled post-training. The model uses the same base model as GLM-5.2, with performance gains coming from additional post-training environments, more diverse tasks, and expanded compute on the existing training stack. Z.ai’s stack includes IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for large-scale asynchronous post-training. GLM-5.3 is designed to perform better on work that resembles real engineering tasks rather than short coding exercises. Its training environments include production-style workflows where the model must diagnose bottlenecks, inspect documentation, use codebases, run experiments, implement changes, and produce measurable improvements. The model improves coding performance across public and private benchmarks, including Terminal Bench 3.0, DeepSWE, Agents’ Last Exam, and Z.ai Code Bench. GLM-5.3 also improves token efficiency, producing stronger agentic coding results than GLM-5.2 while using fewer output tokens in Z.ai’s internal evaluations. The model supports three reasoning effort levels, low, high, and max, and no longer supports disabling thinking. Z.ai recommends max reasoning effort for coding tasks, while applications using disabled thinking must migrate to enabled thinking before switching to GLM-5.3. The release also reports emergent cyber capabilities, including stronger vulnerability discovery and exploitation-chain reasoning, with open-weight release planned after safety evaluation and hardening. By combining scaled post-training, long-context infrastructure, long-horizon reinforcement learning, coding-agent workflows, benchmark improvements, reasoning controls, and ZCode integration, GLM-5.3 helps developers and researchers work on demanding coding and agentic tasks.

Molmo is an advanced suite of multimodal AI models developed by the Allen Institute for AI (Ai2) that aims to bridge the gap between open-source and proprietary technologies, ensuring competitive performance on various academic assessments and evaluations by human users. Unlike many existing multimodal models that rely on synthetic datasets created from proprietary sources, Molmo is solely trained on publicly accessible data, fostering both transparency and reproducibility within the realm of AI research. A key innovation in Molmo's creation is the inclusion of PixMo, a distinctive dataset that features detailed image captions curated by human annotators through speech-based descriptions, complemented by 2D pointing data that allows models to communicate using both natural language and non-verbal cues. This ability enables Molmo to interact with its environment in a more refined way, such as by indicating particular objects within images, which expands its applicability across various domains, including robotics, augmented reality, and interactive user interfaces. Moreover, the strides made by Molmo are poised to redefine standards for future research and development in multimodal AI, opening up new avenues for exploration and application. As the field evolves, the influence of Molmo's innovative approach could inspire similar projects aimed at enhancing human-AI interaction.

Olmo 3 constitutes an extensive series of open models that include versions with 7 billion and 32 billion parameters, delivering outstanding performance in areas such as base functionality, reasoning, instruction, and reinforcement learning, all while ensuring transparency throughout the development process, including access to raw training datasets, intermediate checkpoints, training scripts, extended context support (with a remarkable window of 65,536 tokens), and provenance tools. The backbone of these models is derived from the Dolma 3 dataset, which encompasses about 9 trillion tokens and employs a thoughtful mixture of web content, scientific research, programming code, and comprehensive documents; this meticulous strategy of pre-training, mid-training, and long-context usage results in base models that receive further refinement through supervised fine-tuning, preference optimization, and reinforcement learning with accountable rewards, leading to the emergence of the Think and Instruct versions. Importantly, the 32 billion Think model has earned recognition as the most formidable fully open reasoning model available thus far, showcasing a performance level that closely competes with that of proprietary models in disciplines such as mathematics, programming, and complex reasoning tasks, highlighting a considerable leap forward in the realm of open model innovation. This breakthrough not only emphasizes the capabilities of open-source models but also suggests a promising future where they can effectively rival conventional closed systems across a range of sophisticated applications, potentially reshaping the landscape of artificial intelligence.

Mistral 7B is a cutting-edge language model boasting 7.3 billion parameters, which excels in various benchmarks, even surpassing larger models such as Llama 2 13B. It employs advanced methods like Grouped-Query Attention (GQA) to enhance inference speed and Sliding Window Attention (SWA) to effectively handle extensive sequences. Available under the Apache 2.0 license, Mistral 7B can be deployed across multiple platforms, including local infrastructures and major cloud services. Additionally, a unique variant called Mistral 7B Instruct has demonstrated exceptional abilities in task execution, consistently outperforming rivals like Llama 2 13B Chat in certain applications. This adaptability and performance make Mistral 7B a compelling choice for both developers and researchers seeking efficient solutions. Its innovative features and strong results highlight the model's potential impact on natural language processing projects.

We are excited to unveil the latest version of our open-source large language model, which includes model weights and initial code for the pretrained and fine-tuned Llama language models, ranging from 7 billion to 70 billion parameters. The Llama 2 pretrained models have been crafted using a remarkable 2 trillion tokens and boast double the context length compared to the first iteration, Llama 1. Additionally, the fine-tuned models have been refined through the insights gained from over 1 million human annotations. Llama 2 showcases outstanding performance compared to various other open-source language models across a wide array of external benchmarks, particularly excelling in reasoning, coding abilities, proficiency, and knowledge assessments. For its training, Llama 2 leveraged publicly available online data sources, while the fine-tuned variant, Llama-2-chat, integrates publicly accessible instruction datasets alongside the extensive human annotations mentioned earlier. Our project is backed by a robust coalition of global stakeholders who are passionate about our open approach to AI, including companies that have offered valuable early feedback and are eager to collaborate with us on Llama 2. The enthusiasm surrounding Llama 2 not only highlights its advancements but also marks a significant transformation in the collaborative development and application of AI technologies. This collective effort underscores the potential for innovation that can emerge when the community comes together to share resources and insights.

The NVIDIA Llama Nemotron family includes a range of advanced language models optimized for intricate reasoning tasks and a diverse set of agentic AI functions. These models excel in fields such as sophisticated scientific analysis, complex mathematics, programming, adhering to detailed instructions, and executing tool interactions. Engineered with flexibility in mind, they can be deployed across various environments, from data centers to personal computers, and they incorporate a feature that allows users to toggle reasoning capabilities, which reduces inference costs during simpler tasks. The Llama Nemotron series is tailored to address distinct deployment needs, building on the foundation of Llama models while benefiting from NVIDIA's advanced post-training methodologies. This results in a significant accuracy enhancement of up to 20% over the original models and enables inference speeds that can reach five times faster than other leading open reasoning alternatives. Such impressive efficiency not only allows for tackling more complex reasoning challenges but also enhances decision-making processes and substantially decreases operational costs for enterprises. Furthermore, the Llama Nemotron models stand as a pivotal leap forward in AI technology, making them ideal for organizations eager to incorporate state-of-the-art reasoning capabilities into their operations and strategies.

Qwen2 is a comprehensive array of advanced language models developed by the Qwen team at Alibaba Cloud. This collection includes various models that range from base to instruction-tuned versions, with parameters from 0.5 billion up to an impressive 72 billion, demonstrating both dense configurations and a Mixture-of-Experts architecture. The Qwen2 lineup is designed to surpass many earlier open-weight models, including its predecessor Qwen1.5, while also competing effectively against proprietary models across several benchmarks in domains such as language understanding, text generation, multilingual capabilities, programming, mathematics, and logical reasoning. Additionally, this cutting-edge series is set to significantly influence the artificial intelligence landscape, providing enhanced functionalities that cater to a wide array of applications. As such, the Qwen2 models not only represent a leap in technological advancement but also pave the way for future innovations in the field.

Kimi K2 showcases a groundbreaking series of open-source large language models that employ a mixture-of-experts (MoE) architecture, featuring an impressive total of 1 trillion parameters, with 32 billion parameters activated specifically for enhanced task performance. With the Muon optimizer at its core, this model has been trained on an extensive dataset exceeding 15.5 trillion tokens, and its capabilities are further amplified by MuonClip’s attention-logit clamping mechanism, enabling outstanding performance in advanced knowledge comprehension, logical reasoning, mathematics, programming, and various agentic tasks. Moonshot AI offers two unique configurations: Kimi-K2-Base, which is tailored for research-level fine-tuning, and Kimi-K2-Instruct, designed for immediate use in chat and tool interactions, thus allowing for both customized development and the smooth integration of agentic functionalities. Comparative evaluations reveal that Kimi K2 outperforms many leading open-source models and competes strongly against top proprietary systems, particularly in coding tasks and complex analysis. Additionally, it features an impressive context length of 128 K tokens, compatibility with tool-calling APIs, and support for widely used inference engines, making it a flexible solution for a range of applications. The innovative architecture and features of Kimi K2 not only position it as a notable achievement in artificial intelligence language processing but also as a transformative tool that could redefine the landscape of how language models are utilized in various domains. This advancement indicates a promising future for AI applications, suggesting that Kimi K2 may lead the way in setting new standards for performance and versatility in the industry.

OLMo 2 is a suite of fully open language models developed by the Allen Institute for AI (AI2), designed to provide researchers and developers with straightforward access to training datasets, open-source code, reproducible training methods, and extensive evaluations. These models are trained on a remarkable dataset consisting of up to 5 trillion tokens and are competitive with leading open-weight models such as Llama 3.1, especially in English academic assessments. A significant emphasis of OLMo 2 lies in maintaining training stability, utilizing techniques to reduce loss spikes during prolonged training sessions, and implementing staged training interventions to address capability weaknesses in the later phases of pretraining. Furthermore, the models incorporate advanced post-training methodologies inspired by AI2's Tülu 3, resulting in the creation of OLMo 2-Instruct models. To support continuous enhancements during the development lifecycle, an actionable evaluation framework called the Open Language Modeling Evaluation System (OLMES) has been established, featuring 20 benchmarks that assess vital capabilities. This thorough methodology not only promotes transparency but also actively encourages improvements in the performance of language models, ensuring they remain at the forefront of AI advancements. Ultimately, OLMo 2 aims to empower the research community by providing resources that foster innovation and collaboration in language modeling.

The newest version of the open-source AI framework, which can be customized and utilized across different platforms, is available in several configurations: 1B, 3B, 11B, and 90B, while still offering the option to use Llama 3.1. Llama 3.2 includes a selection of large language models (LLMs) that are pretrained and fine-tuned specifically for multilingual text processing in 1B and 3B sizes, whereas the 11B and 90B models support both text and image inputs, generating text outputs. This latest release empowers users to build highly effective applications that cater to specific requirements. For applications running directly on devices, such as summarizing conversations or managing calendars, the 1B or 3B models are excellent selections. On the other hand, the 11B and 90B models are particularly suited for tasks involving images, allowing users to manipulate existing pictures or glean further insights from images in their surroundings. Ultimately, this broad spectrum of models opens the door for developers to experiment with creative applications across a wide array of fields, enhancing the potential for innovation and impact.

Mistral Large 3 is a frontier-scale open AI model built on a sophisticated Mixture-of-Experts framework that unlocks 41B active parameters per step while maintaining a massive 675B total parameter capacity. This architecture lets the model deliver exceptional reasoning, multilingual mastery, and multimodal understanding at a fraction of the compute cost typically associated with models of this scale. Trained entirely from scratch on 3,000 NVIDIA H200 GPUs, it reaches competitive alignment performance with leading closed models, while achieving best-in-class results among permissively licensed alternatives. Mistral Large 3 includes base and instruction editions, supports images natively, and will soon introduce a reasoning-optimized version capable of even deeper thought chains. Its inference stack has been carefully co-designed with NVIDIA, enabling efficient low-precision execution, optimized MoE kernels, speculative decoding, and smooth long-context handling on Blackwell NVL72 systems and enterprise-grade clusters. Through collaborations with vLLM and Red Hat, developers gain an easy path to run Large 3 on single-node 8×A100 or 8×H100 environments with strong throughput and stability. The model is available across Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Fireworks, OpenRouter, Modal, and more, ensuring turnkey access for development teams. Enterprises can go further with Mistral’s custom-training program, tailoring the model to proprietary data, regulatory workflows, or industry-specific tasks. From agentic applications to multilingual customer automation, creative workflows, edge deployment, and advanced tool-use systems, Mistral Large 3 adapts to a wide range of production scenarios. With this release, Mistral positions the 3-series as a complete family—spanning lightweight edge models to frontier-scale MoE intelligence—while remaining fully open, customizable, and performance-optimized across the stack.

Smaug-72B stands out as a powerful open-source large language model (LLM) with several noteworthy characteristics: Outstanding Performance: It leads the Hugging Face Open LLM leaderboard, surpassing models like GPT-3.5 across various assessments, showcasing its adeptness in understanding, responding to, and producing text that closely mimics human language. Open Source Accessibility: Unlike many premium LLMs, Smaug-72B is available for public use and modification, fostering collaboration and innovation within the artificial intelligence community. Focus on Reasoning and Mathematics: This model is particularly effective in tackling reasoning and mathematical tasks, a strength stemming from targeted fine-tuning techniques employed by its developers at Abacus AI. Based on Qwen-72B: Essentially, it is an enhanced iteration of the robust LLM Qwen-72B, originally released by Alibaba, which contributes to its superior performance. In conclusion, Smaug-72B represents a significant progression in the field of open-source artificial intelligence, serving as a crucial asset for both developers and researchers. Its distinctive capabilities not only elevate its prominence but also play an integral role in the continual advancement of AI technology, inspiring further exploration and development in this dynamic field.

Phi-4-reasoning is a sophisticated transformer model that boasts 14 billion parameters, crafted specifically to address complex reasoning tasks such as mathematics, programming, algorithm design, and strategic decision-making. It achieves this through an extensive supervised fine-tuning process, utilizing curated "teachable" prompts and reasoning examples generated via o3-mini, which allows it to produce detailed reasoning sequences while optimizing computational efficiency during inference. By employing outcome-driven reinforcement learning techniques, Phi-4-reasoning is adept at generating longer reasoning pathways. Its performance is remarkable, exceeding that of much larger open-weight models like DeepSeek-R1-Distill-Llama-70B, and it closely rivals the more comprehensive DeepSeek-R1 model across a range of reasoning tasks. Engineered for environments with constrained computing resources or high latency, this model is refined with synthetic data sourced from DeepSeek-R1, ensuring it provides accurate and methodical solutions to problems. The efficiency with which this model processes intricate tasks makes it an indispensable asset in various computational applications, further enhancing its significance in the field. Its innovative design reflects an ongoing commitment to pushing the boundaries of artificial intelligence capabilities.

Liquid AI's LFM2.5 marks a significant evolution in on-device AI foundation models, designed to optimize efficiency and performance for AI inference across edge devices, including smartphones, laptops, vehicles, IoT systems, and various embedded hardware, all while eliminating reliance on cloud computing. This upgraded version builds on the previous LFM2 framework by significantly increasing the scale of pretraining and enhancing the stages of reinforcement learning, leading to a collection of hybrid models that feature approximately 1.2 billion parameters and successfully balance adherence to instructions, reasoning capabilities, and multimodal functions for real-world applications. The LFM2.5 lineup includes various models, such as Base (for fine-tuning and personalization), Instruct (tailored for general-purpose instruction), Japanese-optimized, Vision-Language, and Audio-Language editions, all carefully designed for swift on-device inference, even under strict memory constraints. Additionally, these models are offered as open-weight alternatives, enabling easy deployment through platforms like llama.cpp, MLX, vLLM, and ONNX, which enhances flexibility for developers. With these advancements, LFM2.5 not only solidifies its position as a powerful solution for a wide range of AI-driven tasks but also demonstrates Liquid AI's commitment to pushing the boundaries of what is possible with on-device technology. The combination of scalability and versatility ensures that developers can harness the full potential of AI in practical, everyday scenarios.

EXAONE Deep is a suite of sophisticated language models developed by LG AI Research, featuring configurations of 2.4 billion, 7.8 billion, and 32 billion parameters. These models are particularly adept at tackling a range of reasoning tasks, excelling in domains like mathematics and programming evaluations. Notably, the 2.4B variant stands out among its peers of comparable size, while the 7.8B model surpasses both open-weight counterparts and the proprietary model OpenAI o1-mini. Additionally, the 32B variant competes strongly with leading open-weight models in the industry. The accompanying repository not only provides comprehensive documentation, including performance metrics and quick-start guides for utilizing EXAONE Deep models with the Transformers library, but also offers in-depth explanations of quantized EXAONE Deep weights structured in AWQ and GGUF formats. Users will also find instructions on how to operate these models locally using tools like llama.cpp and Ollama, thereby broadening their understanding of the EXAONE Deep models' potential and ensuring easier access to their powerful capabilities. This resource aims to empower users by facilitating a deeper engagement with the advanced functionalities of the models.

Code Llama is a sophisticated language model engineered to produce code from text prompts, setting itself apart as a premier choice among publicly available models for coding applications. This groundbreaking model not only enhances productivity for seasoned developers but also supports newcomers in tackling the complexities of learning programming. Its adaptability allows Code Llama to serve as both an effective productivity tool and a pedagogical resource, enabling programmers to develop more efficient and well-documented software. Furthermore, users can generate code alongside natural language explanations by inputting either format, which contributes to its flexibility for various programming tasks. Offered for free for both research and commercial use, Code Llama is based on the Llama 2 architecture and is available in three specific versions: the core Code Llama model, Code Llama - Python designed exclusively for Python development, and Code Llama - Instruct, which is fine-tuned to understand and execute natural language commands accurately. As a result, Code Llama stands out not just for its technical capabilities but also for its accessibility and relevance to diverse coding scenarios.

Laguna XS.2 stands out as Poolside's groundbreaking open-weight coding model, noted for being the lightest and fastest in the Laguna lineup. Equipped with a staggering 33 billion parameters organized in a Mixture of Experts structure, of which 3 billion are active, this model has undergone extensive training in-house utilizing 30 trillion tokens. As the most recent generation model available to the public, it features a second-generation architecture and represents Poolside's first open-weight release, benefiting from lessons learned during the Laguna M.1 training process, which utilized synthetic data and reinforcement learning. Tailored specifically to optimize agentic coding workflows, Laguna XS.2 is exceptional in coding, acting, and rapid iteration, particularly within Poolside's coding agent ecosystem. This model is especially beneficial for developers and teams in need of a lightweight and efficient coding solution, as opposed to more complex frontier systems. Released under the flexible Apache 2.0 license, it enables the community to evaluate, refine, quantize, and build upon its weights, fostering an environment of collaborative development. Ultimately, Laguna XS.2 not only serves as a powerful tool for agentic coding but also promotes creativity and experimentation among its users, allowing for a diverse range of applications and enhancements.

Qwen2.5-Max is a cutting-edge Mixture-of-Experts (MoE) model developed by the Qwen team, trained on a vast dataset of over 20 trillion tokens and improved through techniques such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). It outperforms models like DeepSeek V3 in various evaluations, excelling in benchmarks such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, and also achieving impressive results in tests like MMLU-Pro. Users can access this model via an API on Alibaba Cloud, which facilitates easy integration into various applications, and they can also engage with it directly on Qwen Chat for a more interactive experience. Furthermore, Qwen2.5-Max's advanced features and high performance mark a remarkable step forward in the evolution of AI technology. It not only enhances productivity but also opens new avenues for innovation in the field.

Qwen3.8-27B is an open-weights 27B-class model connected to Alibaba’s Qwen3.8 release, built for developers, researchers, and AI teams that need a capable but more deployable model size. Alibaba’s Qwen3.8 launch described the broader model family as optimized for coding and cowork scenarios, including software development, document processing, data analysis, and professional workflows. Reports state that Alibaba planned to open-source Qwen3.8-Max alongside Qwen3.8-27B, expanding access for developers and researchers. Qwen3.8-27B gives builders a smaller alternative to the 2.4T-parameter Qwen3.8-Max model, which third-party coverage describes as Qwen’s first Max-scale model planned for open weights. The model is well suited for coding assistance, local development, agent testing, workflow automation, data analysis, document understanding, and private AI experimentation. QwenCloud documentation lists Qwen3.8-Max as supporting a 1M context window, thinking, function calling, built-in tools, and structured output, showing the broader Qwen3.8 generation’s focus on advanced agent and application workflows. Qwen3.8-27B is especially useful for teams that want Qwen-family capabilities without the infrastructure demands of Max-scale deployment. Community posts around the release point to active interest in Hugging Face, Unsloth GGUF, Ollama, and local inference use cases. Third-party coverage also notes practical hardware discussions around quantized Qwen3.8-27B deployment, including claims that 4-bit variants can fit more easily on consumer or workstation GPUs. The model can be positioned for organizations that need open AI infrastructure, coding agents, local model evaluation, private deployments, and cost-controlled experimentation. By combining open-weight access, a practical 27B model size, Qwen3.8-era performance ambitions, coding-oriented workflows, and local deployment interest, Qwen3.8-27B gives developers a flexible foundation for building AI products and agents.

This repository presents the research preview for LongLLaMA, an innovative large language model capable of handling extensive contexts, reaching up to 256,000 tokens or potentially even more. Built on the OpenLLaMA framework, LongLLaMA has been fine-tuned using the Focused Transformer (FoT) methodology. The foundational code for this model comes from Code Llama. We are excited to introduce a smaller 3B base version of the LongLLaMA model, which is not instruction-tuned, and it will be released under an open license (Apache 2.0). Accompanying this release is inference code that supports longer contexts, available on Hugging Face. The model's weights are designed to effortlessly integrate with existing systems tailored for shorter contexts, particularly those that accommodate up to 2048 tokens. In addition to these features, we provide evaluation results and comparisons to the original OpenLLaMA models, thus offering a thorough insight into LongLLaMA's effectiveness in managing long-context tasks. This advancement marks a significant step forward in the field of language models, enabling more sophisticated applications and research opportunities.

Fastino functions as an advanced AI platform that focuses on open-weight language models and features the Fastino Fine-Tuning Agent. This cutting-edge agent empowers users to define tasks in straightforward language, which then leads to selecting the right architecture, generating relevant training data, performing training and evaluation, and eventually providing a model customized for specific tasks that is ready for implementation. Through a unified interface, users can both start and revisit fine-tuning initiatives, ensuring that the resulting models align with their requirements and can be integrated into their own systems. The models developed by Fastino are optimized for production-grade performance, typically achieving response times under 50 milliseconds, while also safeguarding user ownership and privacy of the model weights. Impressively, the transition from a simple task description to a fully trained model can take just hours, allowing teams to streamline their specialized deployment processes considerably. In addition, Fastino provides a variety of open-source and open-weight models tailored for numerous specialized AI applications, thereby enhancing user accessibility and adaptability. This comprehensive approach not only simplifies the fine-tuning process but also opens up new possibilities for innovation in AI-driven solutions.

GLM-5 is Z.ai’s most advanced open-source model to date, purpose-built for complex systems engineering, long-horizon planning, and autonomous agent workflows. Building on the foundation of GLM-4.5, it dramatically scales both total parameters and pre-training data while increasing active parameter efficiency. The integration of DeepSeek Sparse Attention allows GLM-5 to maintain strong long-context reasoning capabilities while reducing deployment costs. To improve post-training performance, Z.ai developed slime, an asynchronous reinforcement learning infrastructure that significantly boosts training throughput and iteration speed. As a result, GLM-5 achieves top-tier performance among open-source models across reasoning, coding, and general agent benchmarks. It demonstrates exceptional strength in long-term operational simulations, including leading results on Vending Bench 2, where it manages a year-long simulated business with strong financial outcomes. In coding evaluations such as SWE-bench and Terminal-Bench 2.0, GLM-5 delivers competitive results that narrow the gap with proprietary frontier systems. The model is fully open-sourced under the MIT License and available through Hugging Face, ModelScope, and Z.ai’s developer platforms. Developers can deploy GLM-5 locally using inference frameworks like vLLM and SGLang, including support for non-NVIDIA hardware through optimization and quantization techniques. Through Z.ai, users can access both Chat Mode for fast interactions and Agent Mode for tool-augmented, multi-step task execution. GLM-5 also enables structured document generation, producing ready-to-use .docx, .pdf, and .xlsx files for business and academic workflows. With compatibility across coding agents and cross-application automation frameworks, GLM-5 moves foundation models from conversational assistants toward full-scale work engines.

We are thrilled to introduce Codestral, our first code generation model. This generative AI system, featuring open weights, is designed explicitly for code generation tasks, allowing developers to effortlessly write and interact with code through a single instruction and completion API endpoint. As it gains expertise in both programming languages and English, Codestral is set to enhance the development of advanced AI applications specifically for software engineers. The model is built on a robust foundation that includes a diverse selection of over 80 programming languages, spanning popular choices like Python, Java, C, C++, JavaScript, and Bash, as well as less common languages such as Swift and Fortran. This broad language support guarantees that developers have the tools they need to address a variety of coding challenges and projects. Furthermore, Codestral’s rich language capabilities enable developers to work with confidence across different coding environments, solidifying its role as an essential resource in the programming community. Ultimately, Codestral stands to revolutionize the way developers approach code generation and project execution.

Qwen3.6 is a next-generation large language model developed by Alibaba, designed to deliver advanced reasoning, coding, and multimodal capabilities. It builds on the Qwen3.5 series with a strong emphasis on stability, efficiency, and real-world usability. The model supports multimodal inputs, enabling it to process text, images, and video for more complex analysis and decision-making. One of its key strengths is agentic AI, allowing it to perform multi-step tasks and operate more autonomously in workflows. Qwen3.6 is particularly optimized for coding, capable of handling complex engineering tasks at a repository level rather than just individual functions. It uses a mixture-of-experts architecture, with billions of parameters but only a subset activated during each inference, improving efficiency. The model is available in both open-weight and proprietary versions, giving developers flexibility in deployment and customization. It can be integrated into enterprise systems, APIs, and cloud environments for production use. Qwen3.6 also offers strong multimodal reasoning, enabling it to analyze documents, visuals, and structured data together. It is designed to support a wide range of applications, from software development to data analysis and automation. The model includes enhancements in performance, scalability, and usability compared to earlier versions. It reflects a broader shift toward agent-based AI systems that can execute tasks rather than just provide responses. Overall, Qwen3.6 represents a powerful and versatile AI model for modern enterprise and developer use cases.

The Mixtral 8x7B model represents a cutting-edge sparse mixture of experts (SMoE) architecture that features open weights and is made available under the Apache 2.0 license. This innovative model outperforms Llama 2 70B across a range of benchmarks, while also achieving inference speeds that are sixfold faster. As the premier open-weight model with a versatile licensing structure, Mixtral stands out for its impressive cost-effectiveness and performance metrics. Furthermore, it competes with and frequently exceeds the capabilities of GPT-3.5 in many established benchmarks, underscoring its importance in the AI landscape. Its unique blend of accessibility, rapid processing, and overall effectiveness positions it as an attractive option for developers in search of top-tier AI solutions. Consequently, the Mixtral model not only enhances the current technological landscape but also paves the way for future advancements in AI development.

OpenPipe presents a streamlined platform that empowers developers to refine their models efficiently. This platform consolidates your datasets, models, and evaluations into a single, organized space. Training new models is a breeze, requiring just a simple click to initiate the process. The system meticulously logs all interactions involving LLM requests and responses, facilitating easy access for future reference. You have the capability to generate datasets from the collected data and can simultaneously train multiple base models using the same dataset. Our managed endpoints are optimized to support millions of requests without a hitch. Furthermore, you can craft evaluations and juxtapose the outputs of various models side by side to gain deeper insights. Getting started is straightforward; just replace your existing Python or Javascript OpenAI SDK with an OpenPipe API key. You can enhance the discoverability of your data by implementing custom tags. Interestingly, smaller specialized models prove to be much more economical to run compared to their larger, multipurpose counterparts. Transitioning from prompts to models can now be accomplished in mere minutes rather than taking weeks. Our finely-tuned Mistral and Llama 2 models consistently outperform GPT-4-1106-Turbo while also being more budget-friendly. With a strong emphasis on open-source principles, we offer access to numerous base models that we utilize. When you fine-tune Mistral and Llama 2, you retain full ownership of your weights and have the option to download them whenever necessary. By leveraging OpenPipe's extensive tools and features, you can embrace a new era of model training and deployment, setting the stage for innovation in your projects. This comprehensive approach ensures that developers are well-equipped to tackle the challenges of modern machine learning.

Qwen3.8-Flash-Next is a pioneering open-weight multimodal Mixture-of-Experts architecture that offers an initial look at the design meant for its successor, Qwen4. This model has been expertly crafted to enhance various aspects such as attention mechanisms, residual pathways, embeddings, and optimization strategies, thereby increasing its overall functionality, enhancing computational efficiency, expanding its model capacity, and ensuring stability during training. Its unique hybrid structure combines Gated DeltaNet, which effectively condenses historical information, with Qwen Sparse Attention, facilitating the selection of meaningful context on a micro-block scale to reduce both attention and indexing expenses for lengthy sequences. The Gated Residual feature enhances the residual pathway by incorporating four streams, which helps in dynamically regulating the information flow across different layers. Moreover, the N-gram Embedding cleverly merges large-scale local-pattern memory with minimal computational overhead for each token, with the capability to transfer to host memory for added efficiency. The entire model is built around a main network comprising 125 billion parameters, supplemented by an additional 51 billion parameters specifically for N-gram embeddings, activating only 6 billion parameters for each token processed. This advanced framework underscores the continuous evolution in machine learning architectures, laying the groundwork for exciting future innovations, and it exemplifies the increasing sophistication and potential of multimodal models in various applications.