Модели Deepseek

DeepSeek v4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental v4 Flash Vision Exp.

DeepSeek-V3.2 is a large language model optimized for high computational efficiency and strong tool-use reasoning. It features DeepSeek Sparse Attention (DSA), a mechanism that lowers training and inference costs while maintaining quality in long-context tasks. A scalable reinforcement learning post-training framework further enhances reasoning, achieving performance comparable to GPT-5 and earning top results on the 2025 IMO and IOI. V3.2 also leverages large-scale agentic task synthesis to improve reasoning in practical tool-use scenarios, boosting its generalization and compliance in interactive environments.

DeepSeek v4 Flash 0731 is the official release of DeepSeek v4 Flash, superseding the preview version, with substantially enhanced agentic capabilities. It is a sparse mixture-of-experts model with 13B active parameters out of 284B total, and this re-post-trained revision is suited for coding, reasoning, and agent workflows. The model natively supports a 1M-token context window and flexible reasoning effort: low for simple tasks, high for daily agent workflows, and max for complex ones.

DeepSeek v4 Flash is an efficiency-focused Mixture-of-Experts model from DeepSeek with 284B total parameters and 13B active parameters, supporting a 1M-token context window. It is built for fast inference and high-throughput workloads while preserving strong reasoning and coding capabilities. The model features hybrid attention for efficient long-context processing and offers configurable reasoning modes. It is a strong fit for use cases such as coding assistants, chat applications, and agent workflows where responsiveness and cost efficiency matter.

DeepSeek v4 Pro is a large-scale Mixture-of-Experts model from DeepSeek with 1.6T total parameters and 49B active parameters, supporting a 1M-token context window. It is designed for advanced reasoning, coding, and long-horizon agent workflows, delivering strong results across knowledge, mathematics, and software engineering benchmarks. Built on the same architecture as DeepSeek v4 Flash, it adds a hybrid attention system for efficient long-context processing and supports multiple reasoning modes to balance speed and depth based on the task. It is well suited for demanding workloads such as full-codebase analysis, multi-step automation, and large-scale information synthesis, where both performance and efficiency are essential.

DeepSeek-V3.2-Exp is an experimental large language model from DeepSeek, serving as an intermediate step between V3.1 and future architectures. It features DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that enhances training and inference efficiency for long-context tasks while preserving high output quality.

DeepSeek-V3.1 Terminus is an enhanced version of DeepSeek V3.1 that retains the original model’s capabilities while resolving user-reported issues, such as language consistency and agent functionality. The update further refines the model’s performance in coding and search agent tasks. This large-scale hybrid reasoning model (671B parameters, 37B active) supports both thinking and non-thinking modes. Building on the DeepSeek-V3 foundation, it incorporates a two-phase long-context training approach, allowing for up to 128K tokens, and adopts FP8 microscaling for more efficient inference.

DeepSeek-V3.1 is a 671B-parameter hybrid reasoning model (37B active), supporting both "thinking" and "non-thinking" modes via prompt templates. It extends DeepSeek-V3 with two-phase long-context training (up to 128K tokens) and uses FP8 microscaling for efficient inference. The model excels in tool use, code generation, and reasoning, with performance comparable to DeepSeek-R1 but with faster responses. It supports structured tool calling, code agents, and search agents, making it ideal for research and agentic workflows. Successor to DeepSeek V3-0324, it delivers strong performance across diverse tasks.

DeepSeek Prover V2 is a 671B parameter model, speculated to be geared towards logic and mathematics. Likely an upgrade from DeepSeek-Prover-V1.5, but released without an official announcement or detailed documentation.

DeepSeek V3 is a 685B-parameter, mixture-of-experts model and the latest iteration of the flagship chat model family from the DeepSeek team. Succeeds the previous DeepSeek V3 model and demonstrates strong performance across a variety of tasks.

DeepSeek-R1-0528 is a lightly upgraded release of DeepSeek R1, utilizing more compute and advanced post-training techniques to push its reasoning and inference capabilities to the level of flagship models like O3 and Gemini 2.5 Pro. Excels in math, programming, and logic leaderboards, with a distilled 8B-parameter variant that rivals much larger models on key benchmarks.

DeepSeek v4 Pro 0813 is the official release of DeepSeek v4 Pro, superseding the preview version, with greatly enhanced agentic capabilities and performance improvements that are especially pronounced in production environments. It is built on the DeepSeek v4 Pro (Preview) model structure, with a DSpark speculative decoding module attached. The model natively supports a 1M-token context window and flexible reasoning effort: low for simple tasks, high for daily agent workflows, and max for complex ones.