AI Models | Open-Source LLM Directory

Gemma 4 26B A4B IT is a specialized Mixture-of-Experts (MoE) model designed to bridge the gap between lightweight efficiency and high-parameter reasoning. For developers, the standout feature is its architecture: while the model holds 25.2B total parameters, it only activates approximately 3.8B per token. This allows you to deploy a model with the intelligence profile of a 30B+ parameter dense model while maintaining the low latency and reduced compute costs of a much smaller footprint. It is instruction-tuned for high-precision following, making it ideal for complex RAG pipelines, agentic workflows, and real-time conversational interfaces. Whether you are optimizing for throughput in a production environment or seeking deep reasoning capabilities without the massive VRAM overhead, this model provides a highly efficient scaling path for sophisticated text-based applications.

GLM-5.1 marks a strategic shift from chat-based interaction toward autonomous agentic workflows. While previous iterations focused on short-form instruction following, this model is architected to handle long-horizon reasoning and complex, multi-step coding tasks. For developers, this means moving beyond simple snippet generation to delegating entire development cycles, such as debugging large codebases or implementing multi-file features. With a 204,800 token context window, it maintains high retrieval accuracy across extensive documentation and repository structures, making it a viable backbone for autonomous coding agents. Compared to standard LLMs that struggle with task drift during long sessions, GLM-5.1 is optimized for continuity and independent execution, offering a robust API for integrating sophisticated reasoning into existing CI/CD pipelines and developer tools.

Kimi K2.6 is Moonshot AI’s latest multimodal powerhouse, specifically architected to move beyond simple chat and into autonomous engineering workflows. For developers, the real value lies in its high-context reasoning capabilities and its specialized training for long-horizon coding tasks. Unlike standard LLMs that struggle with architectural consistency, K2.6 is optimized for end-to-end development cycles, spanning languages like Python, Rust, and Go. It excels in scenarios where you need to bridge the gap between high-level UI/UX design and functional code, or when orchestrating complex multi-agent systems to solve multi-step logic problems. With a massive 262k context window, it is built for deep repository analysis and maintaining state across extended debugging sessions. If you are building automated DevOps pipelines, complex software agents, or rapid prototyping tools, K2.6 provides the structural stability and multimodal understanding required for production-grade integration.

Pareto-code is a dynamic routing layer designed to optimize the trade-off between coding performance and latency/cost. Instead of hitting a single static model, this router maintains a tiered shortlist of high-performing LLMs, ranked specifically by their coding percentiles from Artificial Analysis. For developers, this means you aren't overpaying for GPT-4o when a smaller, faster model can handle a simple refactor, yet you aren't sacrificing quality on complex algorithmic tasks. The core mechanism relies on the 'min_coding_score' parameter, allowing you to set a quality threshold between 0 and 1. If a task requires high reasoning, the router escalates to top-tier models; if the task is trivial, it routes to more efficient engines. It integrates seamlessly via the OpenRouter API, making it an ideal middle-layer for building autonomous coding agents or IDE extensions where reliability and cost-efficiency are equally critical.

MiMo-V2.5 is Xiaomi's latest native omnimodal model, engineered specifically to bridge the gap between high-end agentic reasoning and production-scale cost efficiency. For developers building autonomous agents or complex multimodal workflows, this model offers a significant shift in the performance-to-cost ratio, delivering Pro-level capabilities at approximately 50% of the typical inference overhead. Unlike models that rely on bolted-on vision encoders, MiMo-V2.5 features a native architecture that enhances its perception of both static images and temporal video data. This makes it particularly effective for real-time visual reasoning, video analysis, and complex tool-use scenarios where context retention is critical. With a massive 1,050,000 token context window, it is designed to handle extensive documentation or long-form video streams without losing coherence. Whether you are integrating via API for mobile ecosystem automation or building sophisticated vision-language applications, MiMo-V2.5 provides a highly scalable alternative to more expensive, heavyweight multimodal models.

mimo-v2.5-pro is Xiaomi’s high-performance flagship model, specifically engineered for developers working on autonomous agents and complex software engineering workflows. Unlike standard chat models, this iteration is optimized for long-horizon reasoning and multi-step task execution, making it a viable backbone for agentic frameworks. It demonstrates significant proficiency in software development lifecycles, as evidenced by its performance on SWE-bench Pro, and handles extended context requirements with a 1.05M token window. For engineers, this means the ability to ingest entire codebases or massive documentation sets to maintain state across complex debugging or refactoring sessions. While many models struggle with the 'drift' seen in long-form reasoning, mimo-v2.5-pro is tuned to maintain logical consistency through deep-reasoning tasks. It is accessible via API, making it a plug-and-play option for integrating advanced reasoning into existing CI/CD pipelines or automated developer tools.

Hy3-preview is a specialized Mixture-of-Experts (MoE) model from Tencent, engineered specifically for agentic workflows and high-throughput production environments. Unlike standard monolithic LLMs, Hy3 offers a unique architectural advantage through configurable reasoning modes. Developers can toggle between disabled, low, and high reasoning levels, providing a granular way to balance inference latency against complex problem-solving capabilities. This makes it particularly effective for multi-step agent loops where cost and speed are as critical as logic. With a substantial 262k context window, it handles long-form documentation and extensive conversation histories without significant degradation. For teams integrating via API, Hy3 serves as a scalable middle ground between lightweight chat models and heavy-duty reasoning engines, offering a predictable way to scale compute resources based on the specific complexity of the incoming task.

DeepSeek-V4-Flash is a high-throughput Mixture-of-Experts (MoE) model engineered specifically for developers prioritizing low-latency inference without sacrificing reasoning depth. With a massive 284B total parameter architecture, it leverages a sparse activation strategy—utilizing only 13B parameters per token—to deliver a performance-to-cost ratio that challenges much larger, dense models. The standout feature for production environments is the 1M-token context window, making it an ideal candidate for long-form document analysis, massive codebase ingestion, and complex multi-turn agentic workflows. Unlike standard lightweight models that struggle with nuance, V4-Flash maintains high instruction-following accuracy, making it a versatile drop-in replacement for RAG pipelines and automated coding assistants where speed and context density are the primary bottlenecks.

DeepSeek-V4-Pro is a high-performance Mixture-of-Experts (MoE) model engineered for developers who require massive scale without the typical latency overhead of dense architectures. With 1.6 trillion total parameters and 49 billion activated per token, it strikes a sophisticated balance between deep reasoning capabilities and computational efficiency. The standout feature for production environments is the expansive 1M-token context window, making it a viable backbone for complex RAG pipelines, long-form codebase analysis, and multi-document synthesis. Unlike many general-purpose models, V4 Pro shows significant strength in structured logic and advanced programming tasks, positioning it as a direct competitor to top-tier frontier models. For integration, its API-first approach allows for seamless deployment into existing workflows, offering a scalable solution for applications requiring high-density information processing and complex instruction following.

Qwen3.6-27B represents a significant step forward for developers seeking a versatile mid-sized model that balances computational efficiency with multimodal intelligence. Unlike purely text-based LLMs, this dense 27B architecture natively handles text, image, and video inputs, making it a robust engine for complex reasoning tasks that require visual context. For engineering teams, the 262k context window is a standout feature, enabling the processing of massive codebases or long-form video documentation without frequent truncation. While larger models offer higher reasoning ceilings, the 27B parameter count is optimized for high-throughput production environments where latency and cost-per-token are critical KPIs. It is particularly well-suited for building agentic workflows, automated video analysis tools, and sophisticated RAG pipelines that incorporate non-textual data. Integration is straightforward via API, offering a scalable alternative to heavier frontier models when deployment speed and multimodal versatility are the primary requirements.

Qwen3.6-Max-Preview is a frontier-class sparse Mixture-of-Experts (MoE) model designed to bridge the gap between general reasoning and specialized agentic workflows. With a massive parameter scale and a 262k context window, it is specifically tuned for high-density tasks like complex codebase navigation, multi-step tool orchestration, and autonomous software engineering. For developers, the primary value lies in its improved reliability during function calling and its ability to maintain coherence across large-scale documentation ingestion. Unlike dense models that may struggle with latency-to-intelligence ratios, this MoE architecture optimizes for high-throughput reasoning, making it a viable backbone for production-grade AI agents and automated DevOps pipelines. If your stack requires deep integration with external APIs or sophisticated code generation within large repositories, this model offers a significant step up in instruction following and structural accuracy.

For developers building high-throughput applications, Qwen3.6-35B-A3B offers a compelling middle ground between lightweight edge models and massive dense architectures. By utilizing a Sparse Mixture-of-Experts (SMoE) design, it delivers the reasoning capabilities of a much larger model while only activating 3 billion parameters per token. This significantly reduces inference latency and compute costs without sacrificing the nuanced understanding required for complex tasks. The model is natively multimodal, making it a versatile choice for pipelines involving both vision and text. With a massive 262k context window, it excels at long-document processing, codebase analysis, and complex RAG workflows. Unlike monolithic models, this architecture is optimized for efficient scaling, allowing you to maintain high performance in production environments where tokens-per-second and cost-efficiency are critical KPIs. Whether you are integrating via API or fine-tuning for specific domain logic, the efficiency-to-intelligence ratio here is highly competitive for modern AI orchestration.

Qwen3.6 Flash is Alibaba's latest efficient language model optimized for speed and cost-effectiveness. It supports text, image, and video inputs with a massive 1M token context window, making it suitable for complex multimodal applications. The model offers tiered pricing, allowing developers to balance performance and budget based on their specific needs. Compared to previous versions, it delivers faster inference times while maintaining strong reasoning capabilities across multiple languages. Integration is straightforward through standard APIs, and it's particularly well-suited for real-time applications like chatbots, content generation, and data analysis where latency matters. Its multilingual support and long-context handling make it a solid choice for global development teams working on scalable AI products.

Qwen3.5-Plus (April 2026) represents a significant leap in multimodal reasoning for developers building complex, data-heavy applications. Unlike previous iterations that focused primarily on text, this model natively processes text, high-resolution images, and video streams within a single inference pass. For engineers, the standout feature is the 1M token context window, which effectively moves long-form video analysis and massive codebase auditing from a retrieval-augmented generation (RAG) problem to a direct context problem. While many models struggle with temporal consistency in video, Qwen3.5-Plus is optimized for long-sequence multimodal understanding. Integration is handled via standard API protocols, making it a drop-in replacement for developers looking to upgrade from text-only LLMs to sophisticated vision-language agents. It is particularly well-suited for automated visual QA, complex video summarization, and multimodal reasoning tasks where high-fidelity spatial understanding is required.

NVIDIA's Nemotron-3-Nano-Omni is a specialized 30B parameter multimodal model engineered specifically for high-efficiency agentic workflows. Unlike massive general-purpose LLMs, this model is architected as a 'sub-agent' designed to handle perception and context management within larger enterprise systems. It processes text, images, and video, making it an ideal candidate for tasks requiring visual reasoning or multi-modal context extraction before passing structured data to a primary orchestrator. For developers, the standout feature is its ability to act as a lightweight, high-speed sensory layer, reducing latency and token costs in complex RAG or agentic pipelines. While it lacks the broad creative depth of much larger models, its optimization for multimodal input and enterprise-grade reliability makes it a powerful tool for building autonomous systems that need to 'see' and 'understand' environment data in real-time.

Mistral Medium 3.5 is a high-density 128B parameter model engineered for developers who require a balance between reasoning depth and operational efficiency. Unlike smaller, specialized models, this iteration excels in complex instruction-following and multimodal processing, allowing you to feed both text and visual data into a single pipeline. It is specifically optimized for agentic workflows, where the model must maintain logical consistency across multi-step reasoning tasks or complex coding environments. For international teams, the model’s strength lies in its ability to handle high-context tasks within a large 262k window, making it ideal for analyzing extensive documentation or long-form codebases. While it competes with top-tier frontier models, its architectural focus on dense instruction following makes it a highly predictable choice for production-grade automation and sophisticated RAG implementations.

Mistral Medium 3.5 is a high-density 128B parameter model engineered specifically for developers building autonomous agentic workflows and sophisticated reasoning pipelines. Unlike smaller, faster models that struggle with long-range logic, this model strikes a balance between high-level cognitive reasoning and practical latency, making it ideal for complex coding tasks and multi-step instruction following. It features native multimodal capabilities, allowing you to process image inputs alongside text, which expands its utility in visual reasoning and document analysis. With a massive 262,144 token context window, it is built to ingest entire codebases or massive datasets without losing coherence. For teams moving beyond simple chat interfaces into structured tool-use and automated decision-making, this model provides the stability and instruction adherence required for production-grade integration via API.

Grok 4.3:batch is a high-throughput reasoning model designed for developers building autonomous agentic workflows and complex instruction-following pipelines. Unlike standard chat models, this iteration prioritizes logical consistency and factual precision, making it ideal for RAG-heavy applications and multi-step task execution. It features native multimodal capabilities, allowing you to process both text and visual data within a single context window of up to 1,000,000 tokens. For teams scaling production environments, the 'batch' designation indicates an optimized endpoint for non-latency-sensitive, high-volume processing, offering a cost-effective way to handle large-scale data extraction, document analysis, and automated reasoning tasks. If your stack requires a model that can maintain deep coherence across massive datasets without the overhead of real-time conversational latency, this is a highly competitive option for your deployment pipeline.

Grok 4.3 marks a significant shift toward high-fidelity reasoning and autonomous agentic workflows. Unlike standard LLMs that prioritize conversational fluency, this model is architected for precision in instruction-following and complex multi-step logic. For developers building autonomous agents, the model's ability to process multimodal inputs—specifically text and imagery—allows for more sophisticated environmental perception in software automation. The 1M token context window is a standout feature, enabling the ingestion of massive codebases or extensive technical documentation for RAG-based applications without losing structural coherence. While many models struggle with factual drift during long-form reasoning, Grok 4.3 is optimized for high-density information retrieval. If your stack requires a reliable engine for tool-calling, automated debugging, or complex data extraction, this model provides the stability and reasoning depth necessary to move beyond simple chat interfaces into production-grade agentic systems.

Perceptron-mk1 is a high-fidelity vision-language model engineered specifically for temporal reasoning and embodied AI applications. Unlike standard multimodal models that treat video as a sequence of static frames, mk1 is optimized to parse complex motion dynamics and spatial relationships over time. For developers building autonomous agents, robotics controllers, or advanced video analytics pipelines, this model provides the granular visual grounding necessary to translate raw video streams into actionable logic. It excels in tasks requiring long-context visual understanding, such as describing multi-step physical actions or diagnosing causal events in a video sequence. Integration is handled via a standard API, supporting a 32k context window to accommodate extended temporal data. While many VLMs struggle with the 'temporal drift' seen in long video clips, mk1 is architected to maintain high-resolution semantic consistency throughout the input stream, making it a robust choice for real-world deployment in vision-centric workflows.

Grok-build-0.1 is a specialized model engineered specifically for agentic software engineering workflows. Unlike general-purpose LLMs that struggle with long-term architectural consistency, this model is optimized for the high-frequency, iterative loops required by autonomous coding agents. It features a massive 256k context window, making it highly effective for ingesting entire codebases or complex documentation sets to maintain state across multi-step tasks. The model supports multimodal inputs, allowing developers to feed in UI screenshots or architectural diagrams to drive implementation. For integration, it is designed for low-latency interaction, which is critical when deploying agents that must perform rapid reasoning and tool-calling cycles. While general models excel at snippet generation, Grok-build-0.1 is purpose-built for the more demanding lifecycle of autonomous debugging, refactoring, and system-level development.

Qwen3.7-Max represents a significant shift toward agentic intelligence, moving beyond simple chat interfaces to handle complex, multi-step reasoning workflows. For developers building autonomous agents, this model is optimized for high-reliability tool use and structured output, making it a strong candidate for integration into automated DevOps pipelines or sophisticated software engineering assistants. Unlike general-purpose models that struggle with long-context consistency, Qwen3.7-Max leverages its massive 1M token window to maintain coherence across entire codebases or extensive technical documentation. While many models focus on creative prose, this flagship is engineered for precision in coding, logical reasoning, and productivity automation. If your stack requires a model that can act as a reasoning engine rather than just a text predictor, Qwen3.7-Max provides the architectural depth needed for production-grade agentic tasks.

Step-3.7-Flash is a high-efficiency multimodal MoE model designed for developers requiring low-latency reasoning and native vision processing. Unlike traditional dense models, it utilizes a 196B-parameter backbone but only activates approximately 11B parameters per token, striking a balance between high-level intelligence and rapid inference speeds. For developers building real-time applications, this architecture minimizes time-to-first-token without sacrificing the depth needed for complex instruction following. The model excels in multimodal workflows, offering integrated image and video understanding that goes beyond simple captioning to include spatial reasoning and temporal analysis. With a massive 262k context window, it is particularly well-suited for long-document processing, large-scale codebase analysis, and multi-frame video reasoning. If your stack requires a scalable API-driven solution that handles vision-heavy tasks with the efficiency of a smaller model, Step-3.7-Flash provides a competitive alternative to standard lightweight LLMs.

MiniMax-M3:batch is a high-throughput multimodal foundation model designed for developers building complex, long-context applications. Unlike standard chat models, M3 is architected to handle massive context windows—up to 1M tokens—making it a viable engine for deep codebase analysis, extensive document processing, and long-horizon agentic workflows. It natively processes text, image, and video inputs, allowing for sophisticated cross-modal reasoning within a single pipeline. For engineering teams, the 'batch' designation suggests an optimization for non-real-time, high-volume processing tasks where cost-efficiency and throughput are prioritized over instantaneous latency. Whether you are automating complex coding tasks or building autonomous agents that require continuous environmental observation via video, M3 provides the structural depth needed to maintain coherence across extended sequences that typically cause smaller models to lose track.