Alibaba Qwen 3

The Qwen3 235B A22B Thinking 2507 is the premier flagship, reasoning-optimized Mixture-of-Experts (MoE) Causal Language Model in the Qwen3 family, built with 235B total parameters and 22B activated per forward pass, designed for advanced analytical and research-grade workloads. This checkpoint operates exclusively in thinking mode and is trained to generate explicit reasoning traces during inference, enabling deeper multi-step problem solving across mathematics, science, coding, and academic benchmarks. It supports a native 262,144-token context window, allowing sustained reasoning over long documents, complex datasets, and extended multi-source inputs without context degradation.

The Qwen3 235B A22B Instruct 2507 is the premier flagship Mixture-of-Experts (MoE) Causal Language Model in the Qwen3 family, built with 235B total parameters and 22B activated per forward pass. This version is optimized for high-quality instruction following, logical reasoning, and preference-aligned text generation, with substantial improvements in multilingual long-tail knowledge coverage and subjective task alignment. It supports a native 262,144-token context window and can be extended up to approximately 1,010,000 tokens, enabling large-scale document analysis, extended conversational sessions, and complex multi-source workflows.

The Qwen3 32B is the largest dense Causal Language Model in the Qwen3 family, built on a 64-layer architecture with 31.2B non-embedding parameters to support high-complexity reasoning and high-fidelity generation at scale. It provides a native 32,768-token context window and can be extended up to 131,072 tokens using YaRN, enabling robust handling of long-form documents, extended analytical chains, and deeply nested multi-turn interactions. Incorporating Qwen3’s hybrid thinking design, the model balances deliberate multi-step reasoning with efficient general-purpose generation.

The Qwen3 30B A3B Thinking 2507 is a reasoning-focused Mixture-of-Experts (MoE) Causal Language Model designed for high-complexity analytical workloads that require explicit, step-by-step deliberation. This model operates exclusively in thinking mode and is trained to expose its internal reasoning process during generation, making it well-suited for advanced tasks in mathematics, science, coding, and academic problem solving. It supports a native 262,144-token context window, enabling deep reasoning over long documents, large codebases, and multi-source inputs without losing coherence.

The Qwen3 30B A3B Instruct 2507 is an instruction-aligned Mixture-of-Experts (MoE) Causal Language Model built on a 48-layer transformer architecture with 29.9B non-embedding parameters, designed for efficient large-scale generation under sparse activation. It activates approximately 3.3B parameters per forward pass from a total parameter count of 30.5B, enabling efficient inference while retaining strong general-purpose performance. With native support for a 262,144-token context window, it is well-suited for long-document processing, extended conversational contexts, and large-context instruction workloads.

The Qwen3 14B is a large-scale Causal Language Model built on a 40-layer architecture with 13.2B non-embedding parameters, designed for workloads that require deeper reasoning capacity and higher generation fidelity across complex tasks. It supports a native 32,768-token context window and can be extended up to 131,072 tokens using YaRN, enabling effective processing of long-form documents, extended analytical prompts, and multi-stage conversational workflows. Leveraging Qwen3’s hybrid thinking design, the model transitions smoothly between deliberate, multi-step reasoning and efficient general-purpose generation.

The Qwen3 8B is a high-capability Causal Language Model built on a 36-layer architecture with 6.95B non-embedding parameters, designed for workloads that demand stronger reasoning depth and higher output quality at scale. It supports a native 32,768-token context window and can be extended up to 131,072 tokens using YaRN, enabling effective handling of long-form documents, complex analytical prompts, and extended multi-turn interactions. Leveraging Qwen3’s hybrid thinking design, the model adapts seamlessly between deliberate reasoning tasks and efficient general-purpose generation.

The Qwen3 4B Thinking 2507 is a reasoning-specialized Causal Language Model built on a 36-layer architecture with 3.6B non-embedding parameters. This updated model operates only in thinking mode and is optimized to generate explicit reasoning traces as part of its output, making it suitable for complex analytical tasks that require deeper logical decomposition. It supports a native 262,144-token context window, enabling extended deliberation over long documents, large problem states, and multi-stage reasoning workflows.

The Qwen3 4B Instruct 2507 is an instruction-aligned Causal Language Model built on a 36-layer architecture with 3.6B non-embedding parameters, optimized for high-quality instruction following and preference-aligned generation. Unlike the base Qwen3 models, this updated model operates exclusively in non-thinking mode and is tuned to deliver direct, well-structured responses without intermediate reasoning traces. It supports a native 262,144-token context window, enabling effective handling of very long documents, extended conversations, and large-context comprehension workloads.

The Qwen3 4B is a mid-scale Causal Language Model built on a 36-layer architecture with 3.6B non-embedding parameters, positioned as a balanced option for workloads that require stronger reasoning depth without sacrificing serving efficiency. It supports a native 32,768-token context window and can be extended up to 131,072 tokens using YaRN, enabling reliable performance on long-form documents, multi-step analysis, and extended conversational workflows. Incorporating Qwen3’s hybrid thinking design, the model adapts smoothly between deliberate reasoning tasks and fast, general-purpose generation.

The Qwen3 1.7B is a compact yet high-capability Causal Language Model built on a 28-layer architecture with 1.4B non-embedding parameters, offering a strong balance between reasoning quality, throughput, and serving efficiency. It supports a native 32,768-token context window and incorporates Qwen3’s hybrid thinking design, enabling flexible operation across step-by-step reasoning tasks and fast, general-purpose dialogue.

Qwen3 0.6B is a highly efficient causal language model from the Qwen3 family, built on a 28-layer architecture with approximately 0.44B non-embedding parameters. It is designed to deliver strong instruction following, logical reasoning, and multilingual understanding relative to its compact size. The model supports hybrid thinking modes, allowing step-by-step reasoning for complex tasks or fast responses for simpler queries. The model supports a 32K token context window, enabling it to handle extended prompts and multi-step interactions.