DeepSeek

DeepSeek Coder V2 Lite Instruct is a lightweight, efficient version of the Coder V2 model, optimized for fast code generation and instruction following. It retains strong coding and reasoning capabilities while being more resource-efficient.

DeepSeek LLM 67B Chat (V1) is a 67-billion parameter large language model developed by DeepSeek. It is designed for general-purpose conversational AI tasks, demonstrating strong instruction-following and dialogue capabilities as part of the DeepSeek LLM series.

DeepSeek R1 is a reasoning-focused language model from DeepSeek, optimized for complex tasks in mathematics, coding, and logic. It utilizes chain-of-thought reasoning to solve problems step-by-step and is available as an open-weight model.

DeepSeek R1 is a reasoning-focused model from the R1 series, optimized for complex tasks requiring step-by-step thinking. It excels in mathematics, coding, and logical reasoning by leveraging advanced reinforcement learning and chain-of-thought methodologies.

A distilled reasoning model from DeepSeek, based on the Qwen3 8B architecture. It is optimized for mathematical and code reasoning tasks, offering strong performance in a lightweight and efficient package.

A distilled version of the DeepSeek R1 reasoning model, built on the Llama 70B architecture. It inherits strong reasoning and chain-of-thought capabilities from the R1 series while being optimized for efficiency. The model excels at complex problem-solving, code generation, and tasks requiring logical deduction.

A distilled 8B parameter model from the DeepSeek R1 series, optimized for fast and efficient reasoning tasks. It inherits strong reasoning capabilities from larger R1 models while being lightweight and cost-effective.

A distilled version of the DeepSeek R1 reasoning model, based on the Qwen 1.5B architecture. It inherits strong reasoning capabilities from the larger R1 model while being significantly smaller and faster, making it suitable for edge deployment and low-latency applications.

A distilled version of the DeepSeek R1 reasoning model, built upon the Qwen 14B architecture. It aims to deliver strong reasoning and problem-solving capabilities in a more compact and efficient form factor.

A distilled version of the DeepSeek R1 reasoning model, built on the Qwen 32B architecture. It inherits strong chain-of-thought reasoning capabilities from the larger R1 model while offering faster inference speeds and lower computational costs. This model is optimized for efficient deployment without sacrificing core reasoning performance.

DeepSeek V3 is a powerful and cost-effective large language model released in December 2024. It excels in complex reasoning, coding, and long-context understanding (up to 128K tokens), while maintaining high inference speed and low operational costs.

DeepSeek V3 0324 is DeepSeek's flagship model, featuring strong reasoning and coding capabilities. It offers a high-performance, cost-effective solution for complex tasks.

DeepSeek V3.1 is a high-performance, non-reasoning variant of the V3 series, optimized for fast response times and cost-efficiency. It is designed for general-purpose tasks where rapid inference and low latency are prioritized over complex chain-of-thought reasoning.

DeepSeek V3.1 (Reasoning) is a specialized variant of the V3.1 model optimized for complex reasoning tasks. It incorporates an enhanced thinking process to deliver more accurate and logical solutions for problems requiring multi-step analysis.

DeepSeek V3.1 Terminus is a non-reasoning variant of the V3.1 model, optimized for high-speed response and cost-efficiency. It is designed for tasks where rapid output and low latency are prioritized over complex chain-of-thought reasoning.

DeepSeek V3.1 Terminus (Reasoning) is a specialized variant of the V3.1 model optimized for complex reasoning tasks. It likely incorporates advanced chain-of-thought or thinking mechanisms to enhance performance on logic, analysis, and problem-solving challenges.

DeepSeek V3.2 (Non-reasoning) is a general-purpose large language model optimized for fast response and low cost. It excels at code generation and processing long contexts, making it suitable for a wide range of non-reasoning tasks.

DeepSeek V3.2 (Reasoning) is a large language model optimized for complex reasoning tasks. It features an enhanced thinking mode or chain-of-thought capability, excelling in multi-step logical deduction, mathematical problem-solving, and code generation.

This is an experimental version of the DeepSeek V3 series, optimized for general-purpose tasks and fast response times. It maintains strong code generation and long-context processing capabilities while operating in a non-reasoning mode for quicker outputs.

An experimental reasoning-focused variant of the DeepSeek V3.2 model, optimized for complex logical deduction and chain-of-thought tasks. It likely maintains strong coding and general capabilities while enhancing performance on problems requiring multi-step reasoning.

A specialized variant of the DeepSeek V3.2 architecture, optimized for enhanced performance in code generation and complex reasoning tasks. It maintains the series' hallmark of high cost-efficiency and rapid response times.

A fast and efficient model from the DeepSeek V4 family, optimized for low-latency responses and general tasks. It excels in code generation and instruction following, but is not designed for complex reasoning or chain-of-thought tasks.

This is the Flash version of the DeepSeek V4 series, optimized for reasoning tasks with a high-effort mode to enhance complex problem-solving. It delivers enhanced reasoning performance while maintaining fast response speeds.

DeepSeek V4 Flash is a fast-response model optimized for reasoning tasks. It is designed to deliver high-quality reasoning outputs with maximum computational effort while maintaining low latency.

This is the non-reasoning variant of the DeepSeek V4 Pro model. It retains the core capabilities of the V4 Pro series, such as strong coding and general knowledge, but is optimized for faster response times and lower cost by omitting the extended reasoning or chain-of-thought process. It is suitable for applications requiring quick, cost-effective responses.

A high-effort reasoning model from the DeepSeek V4 series, optimized for complex problem-solving and deep analytical tasks. It likely employs extended thinking or chain-of-thought processes to tackle challenging queries in coding, mathematics, and logic.

DeepSeek V4 Pro is a high-compute reasoning model optimized for complex logical, mathematical, and coding tasks. It represents the 'Max Effort' variant, utilizing significant computational resources to maximize accuracy and depth in its reasoning chains.

DeepSeek-Coder-V2 is a powerful code generation model designed for complex programming tasks. It excels in code completion, generation, and understanding across multiple languages, featuring strong reasoning capabilities and support for long-context windows.

DeepSeek-V2-Chat is an open-source chat model built on a Mixture-of-Experts (MoE) architecture with 236B total parameters but only 21B activated per token, achieving high efficiency. It supports a 128K context window and demonstrates strong performance in coding, mathematics, and multilingual tasks.

DeepSeek-V2.5 is an advanced Mixture-of-Experts (MoE) large language model, representing an evolution of the V2 series. It is optimized for strong reasoning and coding capabilities while maintaining efficiency. The model supports a long context window.

DeepSeek-V2.5 is a high-performance Mixture-of-Experts (MoE) model optimized for coding and reasoning tasks. It offers a strong balance of capability and cost-efficiency, supporting long context windows for complex applications.