DeepSeek-R1 is a reinforcement learning (RL) driven inference model that addresses issues of repetitiveness and readability within the model. Prior to RL, DeepSeek-R1 introduced cold start data to further optimize inference performance. It performs comparably to OpenAI-o1 in mathematical, coding, and reasoning tasks, and enhances overall effectiveness through meticulously designed training methods.

DeepSeek-V3 is a mixture of experts (MoE) language model with 671 billion parameters, utilizing multi-head latent attention (MLA) and the DeepSeekMoE architecture, combined with a load balancing strategy that does not rely on auxiliary loss, optimizing inference and training efficiency. Pre-trained on 14.8 trillion high-quality tokens and fine-tuned with supervision and reinforcement learning, DeepSeek-V3 outperforms other open-source models and approaches leading closed-source models in performance.

DeepSeek-R1 is a reinforcement learning (RL) driven inference model that addresses issues of repetitiveness and readability in models. Prior to RL, DeepSeek-R1 introduced cold start data to further optimize inference performance. It performs comparably to OpenAI-o1 in mathematical, coding, and reasoning tasks, and enhances overall effectiveness through carefully designed training methods.

DeepSeek-V3 is a mixed expert (MoE) language model with 671 billion parameters, utilizing multi-head latent attention (MLA) and the DeepSeekMoE architecture, combined with a load balancing strategy without auxiliary loss to optimize inference and training efficiency. Pre-trained on 14.8 trillion high-quality tokens and fine-tuned with supervision and reinforcement learning, DeepSeek-V3 outperforms other open-source models and approaches leading closed-source models.

The DeepSeek-R1 distillation model optimizes inference performance through reinforcement learning and cold-start data, refreshing the benchmark for open-source models across multiple tasks.

DeepSeek-R1-Distill-Qwen-32B is a model obtained through knowledge distillation based on Qwen2.5-32B. This model is fine-tuned using 800,000 selected samples generated by DeepSeek-R1, demonstrating exceptional performance in mathematics, programming, and reasoning across multiple domains. It has achieved excellent results in various benchmark tests, including a 94.3% accuracy rate on MATH-500, showcasing strong mathematical reasoning capabilities.

The DeepSeek-R1 distillation model optimizes inference performance through reinforcement learning and cold-start data, refreshing the benchmark for open-source models across multiple tasks.

deepseek-ai/DeepSeek-R1-Distill-Llama-8B.description

DeepSeek-R1-Distill-Qwen-7B is a model obtained through knowledge distillation based on Qwen2.5-Math-7B. This model is fine-tuned using 800,000 selected samples generated by DeepSeek-R1, demonstrating excellent reasoning capabilities. It has performed outstandingly in multiple benchmark tests, achieving a 92.8% accuracy rate on MATH-500, a 55.5% pass rate on AIME 2024, and a score of 1189 on CodeForces, showcasing strong mathematical and programming abilities as a 7B scale model.

The DeepSeek-R1 distillation model optimizes inference performance through reinforcement learning and cold-start data, refreshing the benchmark for open-source models across multiple tasks.

DeepSeek V2.5 combines the excellent features of previous versions, enhancing general and coding capabilities.

DeepSeek-VL2 is a mixture of experts (MoE) visual language model developed based on DeepSeekMoE-27B, employing a sparsely activated MoE architecture that achieves outstanding performance while activating only 4.5 billion parameters. This model excels in various tasks, including visual question answering, optical character recognition, document/table/chart understanding, and visual localization.

QVQ-72B-Preview is a research-oriented model developed by the Qwen team, focusing on visual reasoning capabilities, with unique advantages in understanding complex scenes and solving visually related mathematical problems.

QwQ-32B-Preview is Qwen's latest experimental research model, focusing on enhancing AI reasoning capabilities. By exploring complex mechanisms such as language mixing and recursive reasoning, its main advantages include strong analytical reasoning, mathematical, and programming abilities. However, it also faces challenges such as language switching issues, reasoning loops, safety considerations, and differences in other capabilities.

Qwen2.5 is a brand new series of large language models designed to optimize the handling of instruction-based tasks.

LoRA/Qwen/Qwen2.5-7B-Instruct.description

Qwen2.5-7B-Instruct is one of the latest large language models released by Alibaba Cloud. This 7B model shows significant improvements in coding and mathematics. It also provides multilingual support, covering over 29 languages, including Chinese and English. The model has made notable advancements in instruction following, understanding structured data, and generating structured outputs, especially JSON.

Qwen2.5 is a brand new series of large language models designed to optimize the handling of instruction-based tasks.

Qwen2.5 is a brand new series of large language models designed to optimize the handling of instruction-based tasks.

A large language model developed by the Alibaba Cloud Tongyi Qianwen team

LoRA/Qwen/Qwen2.5-72B-Instruct.description

Qwen2.5-72B-Instruct is one of the latest large language models released by Alibaba Cloud. This 72B model shows significant improvements in coding and mathematics. It also provides multilingual support, covering over 29 languages, including Chinese and English. The model has made notable advancements in instruction following, understanding structured data, and generating structured outputs, especially JSON.

Qwen2.5 is a new large language model series with enhanced understanding and generation capabilities.

Qwen2.5-Coder-7B-Instruct is the latest version in Alibaba Cloud's series of code-specific large language models. This model significantly enhances code generation, reasoning, and repair capabilities based on Qwen2.5, trained on 55 trillion tokens. It not only improves coding abilities but also maintains advantages in mathematics and general capabilities, providing a more comprehensive foundation for practical applications such as code agents.