Qwen3-30B-A3B-Thinking-2507 — an upgraded and specialized version of Qwen3-30B-A3B fine-tuned exclusively for reasoning tasks. With 30.5B total parameters (3.3B active), 128 experts (8 activated per token), and an extended context length of 262,144, this model stands as the ideal open-source solution among mid-sized models for applications requiring high-quality reasoning—whether for tool usage, agent-like capabilities, or generating well-structured, accurate responses to highly complex user queries.
An updated version of Qwen3-30B-A3B with 30.5 billion total parameters (3.3B active) and an extended context length of 262,144, designed for generating instant and accurate responses without intermediate reasoning steps. An exceptionally efficient dialogue model capable of solving both technical and creative tasks—ideal for use in chatbots.
The model is an image-to-video (I2V) generative diffusion model designed for high-quality video synthesis at 480P and 720P resolutions. It incorporates a Mixture-of-Experts (MoE) architecture with two specialized experts (high-noise and low-noise) to enhance model capacity while maintaining computational efficiency. The model supports both prompt-based and prompt-free video generation.
T2V-A14B model supports generating 5s videos at both 480P and 720P resolutions. Built with a Mixture-of-Experts (MoE) architecture, it delivers outstanding video generation quality. On new benchmark Wan-Bench 2.0, the model surpasses leading commercial models across most key evaluation dimensions.
It is a 5 billion-parameter text-to-video (TI2V) generative model designed for high-definition video generation at 720P resolution (1280×704 or 704×1280) with 24fps. Built using the Wan2.2-VAE architecture, it achieves a compression ratio of 16×16×4, enabling efficient deployment on consumer-grade GPUs like the RTX 4090. The model supports both text-to-video and image-to-video generation within a unified framework.
A hybrid model with 355B parameters, combining advanced reasoning, programming with artifact generation, and agent capabilities within a unified MoE architecture featuring an increased number of hidden layers. At launch, the model ranks 3rd globally in average score across 12 key benchmarks. Particularly impressive are its abilities in generating complete web applications, interactive presentations, and complex code. Users need only describe to the model how the program should function and what outcome they expect.
A high-quality agent-oriented model with 106B parameters, optimized for fast inference and moderate hardware requirements, while retaining key capabilities in hybrid reasoning and overall functionality. At launch, the model ranks 6th globally across 12 key benchmarks, demonstrating exceptional speed and outstanding performance in real-world development scenarios. Developers particularly highlight its effectiveness in frontend code autocompletion and code correction tasks.
The new flagship MoE model Qwen3-235B-A22B in the Qwen 3 series features enhanced "thinking" capabilities and an extended context length of 262K tokens. Operating exclusively in thinking mode, it achieves state-of-the-art performance among leading open and proprietary thinking models, surpassing many well-known brands in mathematical computations, programming, and logical reasoning tasks. An ideal choice for complex research tasks requiring advanced agent and analytical capabilities.
A compact MoE model with an architecture of 30.5B total parameters, of which only 3.3B are activated per token, specifically designed to assist in writing software code. The model features agent-like capabilities, supports a context length of 262,144 tokens, and demonstrates excellent performance at relatively low computational cost. These qualities make it an ideal choice for use as a programming assistant, a QA system within programming education platforms, and for integration into tools featuring code autocompletion.
Alibaba's flagship agent-based programming model featuring a Mixture-of-Experts architecture (480 billion total parameters, 35 billion active parameters) with native support for a 256K-token context. Qwen3-Coder's application scenarios cover the entire spectrum of modern software development—from building interactive web applications to modernizing legacy systems—including autonomous feature development spanning backend APIs, frontend components, and databases.
The updated flagship MoE model Qwen3, with 235B parameters (22B active), features a native context length of 256K and supports 119 languages. In its implementation, developers have abandoned the hybrid mode, so the model supports only non-thinking mode. However, improved refinement enables the model to significantly outperform competitors, delivering exceptional results in mathematics, programming, and logical reasoning. Furthermore, the FP8 version allows industrial-scale deployment with a 50% memory saving.
The first Russian language model with 32 billion parameters and a hybrid reasoning mode, combining revolutionary efficiency in processing the Russian language with the ability for deep analytical thinking to solve tasks of any complexity. The model provides twice the computational resource savings compared to foreign counterparts while delivering superior performance, opening new possibilities for autonomous AI agents.
An enormous MoE model containing 1 trillion parameters. The model is specifically designed for autonomous execution of complex tasks, tool usage, and interaction with external systems. Kimi K2 doesn't simply answer questions—it takes action. It represents a new generation of AI assistants capable of independently planning, executing, and monitoring multi-step processes without constant human involvement. This is precisely why developers recommend using the model in agent-based systems.
A compact multimodal model with a Mixture-of-Experts architecture (28B total parameters, 3B active), capable of processing text, images, and video with a context length of up to 131K tokens. The model employs an innovative heterogeneous MoE architecture with separate experts for text and visual data, enabling efficient processing of multimodal information without compromising text performance. It supports two modes: a standard mode (for fast response) and a reasoning mode (for enhanced analytics on complex tasks).
A large MoE model with 456B parameters, a massive context window of 1,000,000 tokens, and a reasoning budget of 40,000 tokens. Thanks to architectural innovations, the model is more resource-efficient compared to models of similar size, making it highly effective for a wide range of intelligent analysis tasks and agent-based applications.
Powerful reasoning with maximum capabilities and minimal resource consumption. 456B parameters, a context window of 1,000,000 tokens, Lightning Attention — a novel approach to the attention mechanism, and an increased reasoning budget of 80,000 tokens.
This is ultimate performance for tackling the most complex research and product challenges in mathematics, programming, bioinformatics, law, finance, and beyond.
DeepSeek-R1-0528 is the first major update to the popular DeepSeek R1 series, released on May 28, 2025. The developers revised their approach to depth of thought, and the number of parameters increased to 685 billion, resulting in an improvement of more than 10 percentage points across nearly all significant benchmarks compared to the version released on January 22, 2025.
DeepSeek-R1-0528-Qwen3-8B is a compact model based on Qwen3 with 8 billion parameters, distilled from the flagship version DeepSeek-R1-0528. It achieves state-of-the-art (SOTA) results among open-source models in its category. The model is ideally suited for deployment in resource-constrained environments while retaining advanced mathematical and logical reasoning capabilities from the teacher model.
VisualClozePipeline-384 is model for image generation with visual context.
Phi-4-Reasoning is a compact 14-billion-parameter reasoning model that confidently competes with much larger models in mathematics, programming, and scientific tasks. The model is ideally suited for educational and research applications where high-quality logical reasoning is required while efficiently utilizing computational resources.