AI Models | Open-Source LLM Directory

ERNIE-4.5-VL is a high-capacity multimodal Mixture-of-Experts (MoE) model designed for complex reasoning across text and visual domains. For developers, the standout feature is its architectural efficiency: while boasting 424B total parameters, it only activates 47B per token, optimizing inference latency without sacrificing the depth required for high-level cognitive tasks. Unlike standard vision-language models that often treat images as secondary tokens, this model is trained jointly on interleaved data, making it highly effective for document parsing, visual reasoning, and complex scene understanding. It supports a substantial 123,000 token context window, which is critical for analyzing long-form technical documentation or multi-image workflows. While it operates via API, its performance in structured data extraction and multimodal instruction following positions it as a competitive alternative to leading global frontier models, particularly for enterprise-grade applications requiring precise visual-textual alignment.

morph-v3-fast is a specialized inference model engineered specifically for high-speed code transformations. Unlike general-purpose LLMs that attempt to rewrite entire files, this model is optimized for 'apply' operations—taking specific edit snippets and merging them into existing codebases with high precision. It operates at a throughput of approximately 10,500 tokens per second, making it ideal for real-time IDE integrations, automated refactoring pipelines, and large-scale codebase migrations where latency is a critical bottleneck. The model utilizes a strict XML-based prompting schema involving instruction, initial code, and update tags to ensure structural integrity during the merge process. With a 96% accuracy rate on targeted edits and an 8k context window, it functions less like a chatbot and more like a high-performance compiler backend for programmatic code modification.

Morph-v3-large is a specialized application model engineered specifically for high-precision code transformations. Unlike general-purpose LLMs that often struggle with syntax integrity during large-scale refactoring, this model is optimized for 'apply' tasks—taking specific instructions and mapping them onto existing codebases with minimal regression. It operates at an impressive throughput of approximately 4,500 tokens per second, making it viable for real-time IDE integrations or automated CI/CD refactoring pipelines. The architecture supports a massive 262k context window, allowing developers to pass entire modules or complex dependency trees to ensure transformations remain context-aware. Integration requires a structured XML-style prompt format, which enforces a clear separation between logic instructions and the target source code. For teams building automated migration tools, complex linting fixes, or large-scale boilerplate updates, morph-v3-large offers a high-accuracy alternative to standard chat models that frequently hallucinate code changes.

Hunyuan-A13B-Instruct is a high-efficiency Mixture-of-Experts (MoE) model from Tencent, designed to balance massive knowledge capacity with low-latency inference. While it utilizes an 80B total parameter architecture, it only activates 13B parameters per token, making it an ideal candidate for developers needing sophisticated reasoning without the heavy compute overhead of dense large-scale models. A key differentiator is its native support for Chain-of-Thought (CoT) prompting, which significantly improves performance in complex logical reasoning, mathematical problem-solving, and multi-step instruction following. For engineers building agentic workflows or RAG-based systems, the model's ability to process long-context dependencies while maintaining high throughput offers a pragmatic middle ground between lightweight SLMs and massive frontier models. It is best suited for integration into production pipelines where reasoning depth and cost-efficiency are equally critical.

For developers building applications that require high autonomy and minimal interference, dolphin-mistral-24b-venice-edition offers a specialized alternative to standard enterprise models. Built on the Mistral-Small-24B architecture, this fine-tune focuses on removing the restrictive safety guardrails that often trigger false positives in complex reasoning or creative tasks. By utilizing the Dolphin training methodology, it prioritizes instruction adherence and raw capability over pre-programmed refusal patterns. With a 128k context window, it is well-suited for deep document analysis, complex coding assistance, and roleplay scenarios where nuanced, unfiltered responses are critical. While most proprietary models struggle with 'preachiness,' this model provides a predictable, high-fidelity output that respects the developer's prompt intent. It serves as an ideal backbone for local deployments or private API integrations where data sovereignty and unconstrained logic are the primary technical requirements.

Kimi K2 Instruct represents a significant architectural leap for developers seeking high-performance reasoning within a Mixture-of-Experts (MoE) framework. Built by Moonshot AI, the model manages a massive 1-trillion parameter scale, though it maintains efficiency by activating only 32 billion parameters per forward pass. For engineering teams, this means you get the intelligence of a massive model with the lower latency typically associated with much smaller architectures. The model is particularly well-suited for complex multi-step reasoning, sophisticated coding tasks, and long-context information retrieval, supported by a robust 131k context window. Unlike dense models that scale compute linearly with parameter count, K2’s MoE structure allows for more cost-effective scaling and faster inference speeds. If your workflow involves processing large datasets or building autonomous agents that require deep logical consistency, K2 offers a competitive alternative to existing frontier models, providing a highly scalable API for production-grade integration.

For developers building high-scale applications, the qwen3-235b-a22b-2507 model offers a strategic balance between massive parameter capacity and computational efficiency. Utilizing a Mixture-of-Experts (MoE) architecture, it delivers the reasoning depth of a large-scale model while only activating 22B parameters per token. This makes it particularly effective for latency-sensitive workflows like real-time chat agents or complex instruction-following tasks where throughput is critical. With a substantial 262,144 context window, it is well-suited for long-document analysis, codebase reasoning, and multi-turn dialogues that require maintaining deep state. Unlike dense models of similar scale, this MoE approach provides a more cost-effective way to access high-tier intelligence via API, making it a viable backbone for RAG pipelines and sophisticated agentic workflows that demand both multilingual proficiency and high-speed inference.

UI-TARS-1.5-7b is a specialized multimodal agent designed to bridge the gap between LLMs and graphical user interfaces. Unlike general-purpose vision models, this 7B parameter model is fine-tuned specifically for GUI navigation, enabling it to interpret complex desktop, web, and mobile environments with high precision. For developers building autonomous agents or RPA (Robotic Process Automation) tools, this model offers a lightweight yet capable solution for executing click-and-type workflows, navigating non-standard UI components, and even interacting with gaming interfaces. By leveraging reinforcement learning, it moves beyond simple visual description toward actionable decision-making. It is particularly useful for integration into automated testing suites, accessibility tools, or browser-based automation agents where low latency and high spatial reasoning are critical. While smaller than frontier multimodal models, its optimization for pixel-to-action mapping makes it a highly efficient choice for specialized GUI-driven task automation.

Qwen3-Coder-480B-A35B-Instruct is a high-parameter Mixture-of-Experts (MoE) model specifically engineered for complex software engineering workflows. Unlike standard LLMs that focus on simple autocomplete, this model is architected for agentic autonomy. It excels in high-reasoning tasks such as multi-step function calling, tool orchestration, and navigating massive codebases via its extensive 262k context window. For developers building autonomous coding agents or sophisticated IDE extensions, the MoE architecture offers a strategic balance: the intelligence of a massive parameter set with the inference efficiency required for real-time development cycles. It bridges the gap between simple snippet generation and full-scale repository reasoning, making it a strong contender for integration into CI/CD pipelines and automated debugging environments where long-context coherence is non-negotiable.

For developers building high-logic applications, qwen3-235b-a22b-thinking-2507 represents a significant shift toward efficient, large-scale reasoning. Built on a Mixture-of-Experts (MoE) architecture, this model optimizes compute by activating only 22B parameters per token, despite having a massive 235B parameter footprint. This provides the intelligence of a dense flagship model with the latency benefits of a much smaller one. The standout feature is its massive 262k context window, making it a viable candidate for long-form document analysis, codebase auditing, and complex multi-turn agentic workflows. Unlike standard chat models, this iteration is specifically tuned for 'thinking'—meaning it excels at chain-of-thought processes required for mathematical proofs, advanced coding, and structured logical deduction. If you are moving beyond simple RAG into autonomous reasoning agents, this model offers the scale and context depth necessary to maintain coherence over extended operations.

GLM-4.5-Air is a high-efficiency Mixture-of-Experts (MoE) model designed specifically for developers building autonomous agent workflows. While it maintains the reasoning capabilities of the flagship GLM-4.5 series, this 'Air' variant is optimized for lower latency and reduced computational overhead, making it ideal for high-throughput production environments. For developers, the primary value proposition lies in its agent-centric architecture, which excels at tool calling, multi-step planning, and following complex instructions within long-context windows. Unlike heavy-parameter dense models, its MoE structure allows for faster inference speeds without a significant drop in logical reasoning. It is best suited for integration into RAG pipelines, automated coding assistants, and complex task-oriented bots where response time and cost-per-token are critical constraints. If you are transitioning from smaller models to a more capable reasoning engine, this provides a balanced middle ground between raw power and operational efficiency.

GLM-4.5 represents a significant shift toward agentic workflows, moving beyond simple chat completion to focus on complex, multi-step reasoning. Built on a Mixture-of-Experts (MoE) architecture, the model optimizes computational efficiency while maintaining high performance across diverse reasoning tasks. For developers, the standout feature is its native optimization for agent-based applications, making it a strong candidate for autonomous tool-use, complex planning, and long-context retrieval. With a 128k token context window, it handles large-scale documentation and codebase analysis effectively. Compared to previous iterations, GLM-4.5 offers improved instruction following and more reliable structured output, which is critical when integrating LLMs into automated pipelines. Whether you are building sophisticated RAG systems or autonomous software agents, this model provides the stability and reasoning depth required for production-grade deployment via API.

For developers looking to balance high-performance reasoning with low-latency inference, the Qwen3-30B-A3B-Instruct-2507 offers a compelling Mixture-of-Experts (MoE) architecture. While the total parameter count sits at 30.5B, the model only activates approximately 3.3B parameters per token. This design allows for sophisticated instruction following and deep multilingual comprehension without the massive computational overhead typically associated with dense 30B-class models. It is optimized for standard instruction-following tasks rather than extended 'thinking' or chain-of-thought reasoning, making it an ideal candidate for real-time applications like chat interfaces, automated coding assistants, and complex data extraction. With a massive 262,144 context window, it handles long-form document analysis and large codebase ingestion far more effectively than most mid-sized models. Integration is straightforward via API, providing a scalable solution for production environments where throughput and cost-per-token are critical KPIs.

For developers building autonomous agents or complex software tooling, qwen3-coder-30b-a3b-instruct represents a significant shift toward efficient, high-density reasoning. Unlike dense models of similar scale, this 30.5B parameter Mixture-of-Experts (MoE) architecture utilizes only 8 active experts per forward pass, offering a high performance-to-latency ratio that is critical for real-time IDE integrations. The model is specifically tuned for repository-scale context, moving beyond simple snippet completion to handle deep dependency logic and multi-file architectural understanding. It excels in agentic workflows, showing improved reliability in structured tool calling and function execution compared to previous iterations. Whether you are integrating it via API for automated code reviews or deploying it as a local reasoning engine, its ability to navigate massive context windows makes it a viable alternative to much larger, more expensive proprietary models.

Codestral-2508:batch is Mistral's latest specialized iteration optimized for high-throughput coding workflows. Unlike general-purpose LLMs that prioritize conversational fluidity, this model is architected specifically for the technical demands of the SDLC. It excels in low-latency, high-frequency operations, making it an ideal engine for IDE integrations where speed is critical. Developers can leverage its advanced Fill-In-the-Middle (FIM) capabilities for seamless code completion, automated unit test generation, and rapid debugging. With a massive 256,000 token context window, it handles large-scale repository analysis and complex refactoring tasks that would typically exceed standard model limits. While many models struggle with long-range dependencies in large codebases, Codestral provides the structural awareness necessary for maintaining consistency across multiple files. For teams looking to automate CI/CD linting or build sophisticated autonomous coding agents, this model offers a highly efficient, API-driven solution that balances performance with significant operational scale.

Codestral-2508 is Mistral AI's latest specialized release designed specifically for the high-velocity requirements of modern software engineering workflows. Unlike general-purpose LLMs that struggle with the nuances of syntax-heavy tasks, this model is optimized for low-latency execution, making it an ideal candidate for IDE integrations and real-time autocomplete engines. A key technical differentiator is its refined proficiency in Fill-In-the-Middle (FIM) patterns, which allows for much more accurate code insertion and context-aware completions compared to standard causal models. Beyond simple generation, it excels at structural debugging, automated test suite construction, and complex refactoring tasks. With a massive 256k context window, developers can feed entire repository structures into the prompt to maintain architectural consistency. For teams looking to build custom coding assistants or automated CI/CD agents, Codestral-2508 provides a highly efficient, API-driven backbone that prioritizes speed and precision over broad, conversational fluff.

GLM-4.5V is a high-capacity multimodal foundation model designed specifically for developers building vision-centric agentic workflows. Moving beyond simple image captioning, this model leverages a Mixture-of-Experts (MoE) architecture—utilizing 106B total parameters with a highly efficient 12B active parameter count—to balance deep reasoning with inference speed. For developers, the primary value lies in its sophisticated video understanding and spatial reasoning capabilities, making it a strong candidate for complex automation tasks like UI navigation, video analysis, and real-time visual monitoring. Compared to monolithic dense models, its MoE structure offers a more granular approach to processing diverse visual inputs, providing a scalable backbone for applications requiring high-fidelity multimodal integration via API. It is particularly suited for developers looking to bridge the gap between raw visual data and actionable logic in autonomous agent systems.

Mistral Medium 3.1:batch is an optimized, high-throughput iteration of Mistral's enterprise-grade architecture, specifically engineered for large-scale asynchronous processing. For developers managing high-volume workloads, this batch-optimized version offers a strategic middle ground between lightweight models and heavy-duty frontier models. It maintains high reasoning capabilities and complex instruction following while significantly lowering the cost-per-token compared to real-time inference. The 128k context window makes it ideal for processing massive datasets, long-form document summarization, or large-scale data extraction tasks where immediate latency is less critical than cost efficiency and throughput. Integrating this via API allows for seamless scaling of background jobs, such as offline content moderation, batch translation, or bulk analytical labeling, without the overhead of maintaining real-time connection stability for massive payloads.

Mistral Medium 3.1 is a strategic update to the previous Medium iteration, specifically tuned to bridge the gap between cost-efficiency and frontier-level reasoning. For developers building production-grade applications, this model offers a sweet spot: it delivers the high-order logic required for complex instruction following and structured data extraction without the prohibitive latency or inference costs of massive-scale models. With a 128k context window, it is well-suited for long-form document analysis, RAG pipelines, and multi-turn conversational agents. Unlike many general-purpose models that prioritize broad chat capabilities, Medium 3.1 is optimized for enterprise workflows where reliability and predictable output formats are paramount. It integrates seamlessly via API, making it a viable backbone for developers looking to scale sophisticated agentic workflows or automated reasoning tasks while maintaining tight control over operational overhead.

DeepSeek-V3.1 is a high-density hybrid reasoning model designed for developers who need to toggle between rapid inference and deep logical processing. Built on a massive 671B parameter architecture with only 37B active parameters per token, it offers a highly efficient MoE (Mixture-of-Experts) structure that balances throughput with sophisticated reasoning capabilities. The standout feature is its dual-mode execution: you can trigger a 'thinking' mode for complex algorithmic tasks and multi-step logic, or use standard non-thinking modes for low-latency text generation and chat applications. With a 164k context window, it is well-suited for large-scale codebase analysis, long-document summarization, and complex RAG pipelines. For integration, the model provides a predictable API surface that allows you to control reasoning depth via specific prompt templates, making it a versatile alternative to closed-source frontier models when optimizing for both cost and intelligence.

Hermes-4-405B is a high-parameter reasoning model engineered by Nous Research, leveraging the Meta-Llama-3.1-405B architecture as its foundation. Unlike standard LLMs that provide immediate token streams, this model implements a hybrid reasoning mode. It can autonomously decide when to engage in internal deliberation before delivering a final response, making it particularly effective for complex logic, multi-step mathematical problems, and deep code synthesis. For developers, this means a significant reduction in hallucination rates for high-stakes reasoning tasks. The model supports a massive 131k context window, allowing for the ingestion of extensive documentation or entire codebases. While it carries the raw power of a 405B parameter model, the integrated reasoning capability offers a more efficient alternative to manual chain-of-thought prompting. It is designed for seamless API integration into agentic workflows where decision-making accuracy is more critical than raw throughput.

For developers building agentic workflows or complex reasoning pipelines, qwen3-30b-a3b-thinking-2507 represents a significant step in specialized MoE architectures. Unlike standard dense models, this 30B parameter Mixture-of-Experts model is purpose-built for deep reasoning tasks where accuracy in multi-step logic is more critical than raw token throughput. The standout feature is its dedicated 'thinking mode,' which isolates internal reasoning traces from the final output. This architectural choice is a game-changer for debugging and observability, allowing you to inspect the model's chain-of-thought without polluting your application's primary response stream. While it may not match the sheer speed of smaller, general-purpose models, its ability to handle intricate instruction following and mathematical or logical decomposition makes it a superior choice for RAG-based reasoning, code generation, and automated problem-solving agents. It integrates seamlessly via API, providing a high-intelligence backbone for developers who need verifiable logic rather than just probabilistic text completion.

Kimi-k2-0905 is the latest iteration of Moonshot AI’s large-scale Mixture-of-Experts (MoE) architecture, specifically optimized for high-throughput reasoning and complex instruction following. For developers, the core value lies in its massive 1-trillion parameter scale, which enables sophisticated pattern recognition and deep logical reasoning across diverse datasets. Unlike dense models, its MoE structure offers a more efficient compute-to-performance ratio, making it a strong candidate for agentic workflows and long-context reasoning tasks. The model supports a significant context window of 262,144 tokens, allowing for the processing of entire codebases or extensive documentation in a single pass. Whether you are building autonomous agents, complex RAG pipelines, or advanced coding assistants, Kimi-k2-0905 provides the architectural depth required for production-grade applications where precision and context retention are non-negotiable.

Qwen-plus-2025-07-28 is a high-throughput reasoning model built on the Qwen3 architecture, specifically designed for developers needing a middle ground between lightweight chat models and heavy-duty reasoning engines. The standout feature is its 1-million-token context window, which makes it highly effective for long-form document analysis, large-scale codebase ingestion, and complex multi-turn retrieval tasks. Unlike ultra-large models that trade latency for intelligence, this model optimizes for a 'balanced' profile—providing significant reasoning depth while maintaining the speed and cost-efficiency required for production-scale RAG pipelines and automated agent workflows. For developers integrating via API, it offers a predictable performance-to-cost ratio, making it an ideal candidate for scaling enterprise applications that require both deep context handling and rapid inference cycles.