Alternatives to LFM3

Most AI video tools hand you a black box: closed weights, a subscription, and no way to see what is happening under the hood. LTX takes the opposite approach. Built by Lightricks, LTX is an open foundation model that generates and simulates across video, audio, and the physical world, and it puts the weights, the code, and the control in your hands. At the center of the model is LTX-2.5, a 22B-parameter dual-stream diffusion transformer that produces native 4K video at up to 50 frames per second, with audio and video generated together in a single pass rather than stitched together afterward. Artificial Analysis, an independent benchmarking group, currently ranks LTX among the top three AI video models in the world. You choose how you want to use it. Download the open weights and run LTX-2.5 on your own hardware. License the model for on-premise deployment backed by enterprise support. Or build directly on LTX Studio, the production suite that turns the model into a full creative workflow. Companies like ElevenLabs, Asteria Film Co., Magnopus, and NVIDIA already rely on LTX for their own work. LTX is not built for one-off social clips. It is infrastructure for teams that generate motion, audio, and physical environments as part of their own products and pipelines.

NVIDIA Nemotron 3.5 Bolt is a series of efficient AI models developed by NVIDIA for running autonomous agents and conversational applications directly on local hardware. The family consists of Bolt 74B-A9B 24GB and Bolt E4B 4GB, which address different performance and memory requirements. Bolt 74B-A9B uses a mixture-of-experts architecture with 74 billion total parameters and approximately 9 billion active parameters to balance reasoning capabilities with computational efficiency. Native 2-bit quantization enables the model to operate within a 24GB memory footprint on supported NVIDIA systems. Its agentic capabilities include web research, code generation, software debugging, information retrieval, and multistep task execution. The model can be integrated with local AI agent frameworks and development tools to complete assignments without sending application workloads to external inference services. Bolt E4B is optimized for fast conversational responses, multistep interactions, and embedded AI experiences such as virtual assistants and interactive game characters. NVIDIA supports local deployment of the Bolt family across compatible GeForce RTX, RTX PRO, DGX Spark, and Jetson hardware, depending on the model variant. Bolt 74B-A9B is available with open weights through Hugging Face and supports inference tools including llama.cpp, Ollama, LM Studio, Unsloth, and vLLM.

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.

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.

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.

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-Flash is Xiaomi MiMo’s efficiency-focused open-source omnimodal model for coding, automation, visual work, and agentic applications. It is designed to provide a balance between model capability, inference cost, and practical performance across a broad range of workloads. The model can perform software engineering tasks, use tools, execute multi-step workflows, and interact with computer environments. Its multimodal capabilities support applications such as frontend development, presentation design, 3D content creation, game development, and visual reasoning. MiMo-V2.6 can also use multi-view visual inputs in embodied simulation environments to reason about scenes and guide actions through feedback loops. Xiaomi trained the Flash model using reinforcement learning over roughly 750,000 trajectories spanning coding, general agents, visual tasks, and cybersecurity environments. During that training process, Xiaomi reports substantial gains in long-horizon software engineering and general workflow performance compared with the model’s earlier checkpoints. The company has open-sourced the broader MiMo-V2.6 release along with its technical report, reinforcement learning environments, and RL code to support research and reproducibility. MiMo-V2.6-Flash can be accessed through MiMo Desktop, AI Studio, MiMo Code, the MiMo API Platform, OpenRouter, and Xiaomi MiMo’s open-source distribution channels.

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.

GLM-5.3-Flash is a multimodal foundation model from Z.ai built for high-efficiency reasoning, coding, agents, and visual understanding. The model contains 320 billion parameters in total but activates only 18 billion parameters during inference, helping reduce compute requirements. Its architecture combines linear attention with sparse attention so it can efficiently handle both local dependencies and relevant information spread across very long contexts. Z.ai also introduced IndexPool to reduce the memory and latency overhead associated with long-context retrieval at context lengths reaching one million tokens. The model was pretrained on a 30-trillion-token multimodal dataset that incorporates both textual and visual information. GLM-5.3-Flash is designed for software engineering tasks, autonomous workflows, frontend development, computer use, document analysis, and other professional workloads that benefit from visual reasoning. Its visual coding capabilities allow it to inspect rendered interfaces, identify layout or interaction problems, and use those observations to revise its work. Benchmark results published by Z.ai show that it improves substantially over GLM-5.2 on multiple coding and agentic tests while remaining competitive with more expensive frontier models. GLM-5.3-Flash can be accessed through Z.ai services and is also available as downloadable model weights for deployment through supported open inference frameworks.

MiniMax M3 is a frontier open-weight AI model built for coding, agentic work, multimodal understanding, and ultra-long-context tasks. The model supports up to a 1 million token context window, allowing it to work across large codebases, long documents, logs, project histories, and complex task environments. MiniMax M3 introduces MiniMax Sparse Attention, a sparse attention architecture designed to make long-context processing more efficient. The model is natively multimodal, with training that supports deeper semantic fusion across text, image, and video inputs. It is designed to support software engineering tasks, repository analysis, terminal-style work, browser-style retrieval, tool use, and autonomous workflows. MiniMax M3 has a mixture-of-experts architecture with hundreds of billions of total parameters and a smaller activated parameter count for more efficient inference. Developers can use it for AI coding assistants, workflow automation, research agents, document analysis, visual reasoning, and enterprise AI systems. Its long-context capability makes it especially useful when tasks require many files, references, instructions, or interaction histories to stay available at once. MiniMax M3 helps teams build more capable AI agents that can understand larger problems, work across multiple modalities, and execute complex tasks with stronger context awareness.

Gemma 4 is an advanced AI model developed by Google as part of its Gemini architecture, designed to deliver strong performance while remaining accessible to developers. The model is optimized to run on a single GPU or TPU, allowing more organizations and researchers to experiment with powerful AI technology. Gemma 4 improves natural language understanding and generation, making it suitable for applications such as chatbots, text analysis, and automated content creation. Its architecture enables the model to process complex language patterns while maintaining efficient computational performance. Developers can integrate Gemma 4 into various AI projects that require intelligent text processing or conversational capabilities. The model is designed with scalability in mind, allowing it to support both research experiments and production systems. By offering high-performance AI in a more accessible format, Gemma 4 lowers the barrier for developing sophisticated AI solutions. Its flexibility makes it useful for industries ranging from technology and education to business automation. Researchers can also use the model to explore new AI techniques and improve language processing systems. Overall, Gemma 4 represents a step forward in making powerful AI models easier to deploy and use.

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.

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.

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.

DeepSeek-V4.1-Flash-1017 is an open-weight multimodal AI model designed for complex reasoning, autonomous agents, software development, and visual analysis. The model is an updated version of DeepSeek-V4.1-Flash, with improvements in benchmark performance, execution speed, and output efficiency. It supports text and image processing, enabling applications that combine natural language understanding with visual interpretation and reasoning. Its one-million-token context window allows developers to analyze large documents, extensive software repositories, and lengthy interactions while maintaining contextual information. Coding capabilities include generating software, debugging existing applications, modifying codebases, and completing multistep engineering assignments. The model also supports end-to-end project generation, including interactive games, 3D scenes, and other software applications. Agentic functionality includes external tool calling, configurable reasoning effort, multi-turn conversations, and workflows involving automated research and task execution. DeepSeek provides prompt encoding libraries, inference examples, and compatibility with deployment frameworks such as Hugging Face Transformers, vLLM, and SGLang. DeepSeek-V4.1-Flash-1017 is distributed under the MIT License and can be accessed through DeepSeek's website, mobile application, and API or deployed using its publicly available model weights.

Muse Glimmer is an open-weights model featuring 30 billion parameters, developed by Meta Superintelligence Labs, and is fine-tuned for continuous local agent operations. Its compact design allows it to function on a standard Mac or PC equipped with a single consumer GPU, making it ideal for various tasks such as local agent management, function calling, programming, and LLM-as-a-judge evaluations without reliance on cloud services or internet connectivity. This innovative model integrates advanced capabilities such as long-horizon execution, accurate tool invocation, multimodal comprehension, extended memory for context, and effective instruction following. It is proficient in accomplishing end-to-end tasks as an agent, maintains the ability to engage in multi-step reasoning over lengthy processes, can recover gracefully from failed or unanticipated tool engagements, and interprets interleaved text and images using a specialized perception encoder designed for analyzing screenshots, graphs, and document files. Furthermore, Muse Glimmer is compatible with OpenClaw and other orchestration frameworks, allowing for adjustable reasoning efforts, and has been developed with a diverse dataset encompassing over 100 languages. The model's versatility ensures that it can adapt to various applications, thus enhancing its utility in different domains.

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.

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.

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.

Pixtral Large is an expansive multimodal model featuring 124 billion parameters, crafted by Mistral AI and enhancing their previous Mistral Large 2 framework. This model combines a 123-billion-parameter multimodal decoder with a 1-billion-parameter vision encoder, allowing it to excel in the interpretation of various content types, including documents, charts, and natural images, all while retaining superior text comprehension abilities. With the capability to manage a context window of 128,000 tokens, Pixtral Large can efficiently analyze at least 30 high-resolution images at once. It has achieved remarkable results on benchmarks like MathVista, DocVQA, and VQAv2, outpacing competitors such as GPT-4o and Gemini-1.5 Pro. Available for research and educational purposes under the Mistral Research License, it also has a Mistral Commercial License for business applications. This versatility makes Pixtral Large a valuable tool for both academic research and commercial innovations.

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.

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

Phonon serves as an open-weight text-to-speech (TTS) model and inference runtime designed for developers seeking to incorporate speech generation into applications that require local and offline audio capabilities. This innovative system transforms text into spoken audio and has the ability to stream speech in real-time as it is produced. The runtime offers interfaces for Rust and Python, a Swift package for iOS and macOS, JavaScript support for web browsers, and libraries for Android; it can also function as a self-hosted server compatible with OpenAI's API. Depending on the variant of the model used, it accommodates languages such as English, French, German, Spanish, and Portuguese. Phonon is tailored for developers aiming to integrate speech functionalities into mobile, desktop, web, embedded, and server applications, particularly in scenarios requiring offline access or enhanced privacy. The English model, which consists of 31 million parameters, is licensed under Apache 2.0, while the language-specific and multilingual models, featuring 40 million parameters, operate under CC BY-NC 4.0. Additionally, the runtime’s code is available under either the Apache 2.0 or MIT licenses, making it accessible for various development needs. This flexibility and wide-ranging support make Phonon a versatile choice for speech synthesis across different platforms.

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.

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.

FLUX 3 Action is a versatile 7B world-action model aimed at enhancing action prediction for robotics and other environments with high latency demands. This model is built upon the multimodal FLUX 3 backbone and has undergone extensive pretraining on vast datasets encompassing images, videos, and audio, with a notable focus on video content. It then goes through a phase of joint video-action training and fine-tuning tailored to specific robotic applications and action frameworks. By utilizing instructions, visual input from cameras, and robot joint angles, FLUX 3 Action is capable of predicting motor commands alongside anticipated visual outcomes, enabling robots to perform actions, reassess their surroundings, and iterate on their plans. This approach contrasts with methods that treat visual prediction and control as separate processes, as FLUX 3 Action integrates future video predictions with action commands, effectively leveraging knowledge gained from extensive video pretraining to refine robot control mechanisms. Impressively, its single-step 7B model achieves a success rate of 38.3% on the RoboLab-120 benchmark, showcasing its effectiveness in real-world applications. Furthermore, this innovative integration of action and perception marks a significant advancement in the field of robotic control.

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.

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.

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.

Beam represents Reflection’s inaugural open-weight model, characterized as a sparse Mixture-of-Experts framework featuring a staggering 501 billion parameters, of which 23 billion are actively utilized, specifically crafted for tasks involving coding, reasoning, and agentic functions. Its prowess is derived from extensive pretraining and reinforcement learning, having been developed using a vast dataset of 23.8 trillion diverse, high-quality tokens sourced from the web, public domains, and proprietary licensed materials. With a targeted emphasis on enhancing coding and agentic capabilities, Beam is engineered to provide competitive open-weight performance while ensuring efficient inference compute. This model adeptly handles a wide array of tasks, including intricate software engineering, terminal operations, STEM activities, web searches, tool utilization, and general knowledge inquiries. Reinforcement learning techniques have been employed to bolster its abilities in multi-step reasoning, tool application, and responsiveness to environmental feedback. Moreover, users have the flexibility to manage the balance between efficiency and performance through an adjustable reasoning effort parameter, thus tailoring the model's output to better suit their specific needs.

Shieldstral is an innovative multimodal safety classifier with a 3B parameter open-weight structure, adept at assessing text, images, and combined text-plus-image content based on dynamically defined policies during inference. Rather than adhering to a static set of harm categories, it approaches moderation as a binary question-and-answer format: users submit a contextual instruction outlining the evaluation criteria and strictness, pose a yes-or-no safety inquiry, and present the content for assessment. The model processes the “yes” and “no” logits to generate a continuous, calibrated safety score, enabling applications to prioritize or rank outcomes based on confidence levels instead of relying on a single categorical label. This design effectively integrates prompt classification, response moderation, refusal detection, toxicity assessment, and multimodal safety evaluation into a singular interface, empowering teams to modify policies without the need for model retraining. Shieldstral's versatility allows it to analyze prompts, responses, pairs of prompts and responses, images, and images paired with text, making it a comprehensive tool for safety evaluation. As such, it represents a significant advancement in the field of content moderation.

Antares represents a suite of open-weight security small language models specifically designed to identify existing vulnerabilities within extensive codebases. With models like Antares-350M and Antares-1B, organizations can operate them locally or on-site, allowing for the protection of proprietary source code while also minimizing both inference costs and runtime. The process begins with a description of the vulnerability, an advisory, or a CWE category, where the model engages in a step-by-step investigation akin to that of a human analyst, systematically searching for pertinent code patterns, examining potential files, assimilating new information, and altering its approach when certain avenues prove unfruitful. This strategy enables the model to focus its efforts on the files that are most likely to harbor the identified weaknesses. Ultimately, Antares generates a prioritized list of potentially vulnerable source files along with the detailed exploration trail that led to these findings, facilitating easier review and prioritization for teams. Moreover, this capability not only streamlines the vulnerability assessment process but also enhances the overall security posture of the development environment.

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.

Ideogram 4.0 represents a cutting-edge open image model designed for advanced design capabilities, featuring open weights, support for multiple languages, precise layout management, customizable elements, and high-quality 2K imagery. This innovative model caters to developers and businesses aiming to create, refine, and deploy visual intelligence on their own systems. The training methodology for Ideogram 4.0 employs a describe-to-structure-to-recreate process, which involves interpreting scenes, backgrounds, text, and objects as structured data before reconstructing images based on that understanding. This technique enhances the model's grasp of composition, thereby granting teams greater authority over layout, object placement, typography, and overall visual organization. Tailored for practical design applications, it excels in areas such as branding, advertising, fashion, marketing, culinary arts, apparel, social media, photography, and illustration. Since its inception, Ideogram has pioneered text rendering, and version 4.0 introduces bounding-box layout control to ensure that headlines remain easily legible, thus further enhancing its usability in professional settings. Consequently, ideators can leverage this model to streamline their creative processes and achieve remarkable results.

Kimi K2.5 is a powerful multimodal AI model built to handle complex reasoning, coding, and visual understanding at scale. It supports both text and image or video inputs, enabling developers to build applications that go beyond traditional language-only models. As Kimi’s most advanced model to date, it delivers open-source state-of-the-art performance across agent tasks, software development, and general intelligence benchmarks. The model supports an ultra-long 256K context window, making it ideal for large codebases, long documents, and multi-turn conversations. Kimi K2.5 includes a long-thinking mode that excels at logical reasoning, mathematics, and structured problem solving. It integrates seamlessly with existing workflows through full compatibility with the OpenAI SDK and API format. Developers can use Kimi K2.5 for chat, tool calling, file-based Q&A, and multimodal analysis. Built-in support for streaming, partial mode, and web search expands its flexibility. With predictable pricing and enterprise-ready capabilities, Kimi K2.5 is designed for scalable AI development.

The NVIDIA Nemotron 3 Nano Omni represents a groundbreaking open foundation model that integrates various modes of perception and reasoning—including text, images, audio, video, and documents—into a single streamlined architecture. By eliminating the necessity for distinct models tailored to each modality, it effectively minimizes inference delays, simplifies orchestration, and lowers costs while ensuring a cohesive cross-modal context. This innovative model is specifically engineered for agentic AI systems, functioning as a perception and context sub-agent that empowers larger AI entities to perceive and interpret their surroundings in real-time across various formats such as screens, recordings, and both structured and unstructured data. Its capabilities extend to complex multimodal reasoning tasks, encompassing document comprehension, speech recognition, extensive audio-video analysis, and intricate computer workflows, thus allowing agents to navigate dynamic interfaces and multifaceted environments with ease. With a hybrid architecture that is finely tuned for handling long contexts and high throughput, the Nemotron 3 Nano Omni is adept at managing sizable inputs, including multi-page documents, making it a versatile tool in the realm of AI development. Not only does it unify modalities, but it also enhances the overall efficiency of intelligent systems in processing and understanding diverse data types.

Mistral Large 4 serves as a versatile multimodal language model, accessible via Mistral’s hosted inference API. It integrates capabilities for instruction adherence, reasoning, and agent-like functions, allowing it to process both images and text simultaneously. This model is designed for developers to facilitate various tasks such as chat completions, structured data outputs, executing functions, answering questions from documents, and enhancing tool-assisted workflows. Its range of applications spans conversational AI, software development assistance, cybersecurity evaluations, financial and legal processes, scientific inquiry, data extraction, and the interpretation of various document types, including charts and images. Aimed at developers, researchers, and organizations within the public sector, Mistral Large 4 is tailored for creating AI-driven applications across numerous fields, including engineering, manufacturing, finance, logistics, and scientific research. Furthermore, its adaptability makes it suitable for both small startups and large enterprises looking to leverage AI technologies effectively.

Ling 3.0 Tiny is a reasoning model featuring open weights, comprising 7.9 billion total parameters and 1.3 billion active parameters, alongside a substantial context window of 262,000 tokens. Leveraging a mixture-of-experts architecture, it pushes the boundaries of the open-weights Pareto frontier in terms of intelligence relative to active parameters, while being compact enough for local deployment in various environments. Scoring 25 on the Artificial Analysis Intelligence Index, it stands on par with gpt-oss-120b, which scores 24, despite utilizing 15 times fewer total parameters and 4 times fewer active parameters. This impressive parameter efficiency does come with a trade-off, as it requires a significant 213 million output tokens to complete the Intelligence Index evaluation. In addition, Ling 3.0 Tiny exhibits noteworthy advancements in reducing hallucination tendencies compared to Ling-mini-2.0; it enhances its AA-Omniscience score by 59 points while keeping accuracy levels consistent. Notably, rather than making random guesses in uncertain situations, the model chose to attempt only 37% of the questions during evaluation, leading to a markedly reduced hallucination rate of 30%, a significant improvement over the previous generation's 96%. This strategic approach not only demonstrates the model's improved reasoning capabilities but also highlights its potential for more reliable real-world applications.

GLM-4.5V-Flash is a vision-language model that is open source and specifically crafted to integrate robust multimodal functionalities into a compact and easily deployable framework. It accommodates various types of inputs including images, videos, documents, and graphical user interfaces, facilitating a range of tasks such as understanding scenes, parsing charts and documents, reading screens, and analyzing multiple images. In contrast to its larger counterparts, GLM-4.5V-Flash maintains a smaller footprint while still embodying essential visual language model features such as visual reasoning, video comprehension, handling GUI tasks, and parsing complex documents. This model can be utilized within “GUI agent” workflows, allowing it to interpret screenshots or desktop captures, identify icons or UI components, and assist with both automated desktop and web tasks. While it may not achieve the performance enhancements seen in the largest models, GLM-4.5V-Flash is highly adaptable for practical multimodal applications where efficiency, reduced resource requirements, and extensive modality support are key considerations. Its design ensures that users can harness powerful functionalities without sacrificing speed or accessibility.

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.

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.

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.

Qwen3.6-27B is an open-source, dense multimodal language model from the Qwen3.6 series, engineered to provide top-tier performance in areas such as coding, reasoning, and agent-driven workflows, all while maintaining an efficient parameter count of 27 billion. This model is recognized for its ability to outperform or compete closely with much larger counterparts on essential benchmarks, particularly excelling in agent-based coding tasks. It features dual operational modes—thinking and non-thinking—that enable it to effectively adapt its reasoning depth and response speed based on the specific requirements of each task. Additionally, it supports a variety of input types, including text, images, and video, showcasing its versatility. As part of the Qwen3.6 lineup, this model prioritizes practical usability, consistency, and the enhancement of developer productivity, reflecting advancements inspired by community insights and real-world application demands. Its innovative design not only responds to immediate user needs but also anticipates future trends in AI development.