DeepSeek V4
DeepSeek V4 is a large-scale language model released by DeepSeek, employing a Mixture-of-Experts (MoE) architecture with sparse computation characteristics that activate only a subset of parameters during inference. The model demonstrates strong performance in programming, mathematical reasoning, and agent tasks, and belongs to the open-source large language model series, supporting a range of general natural language processing and generation tasks.
Core Facts
DeepSeek V4 adopts a proprietary hierarchical compressed attention strategy, featuring two alternating attention modes: CSA (Compressed Sink Attention) and HCA (Heavy Compressed Attention).
DeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
Timeline (last 90 days)
For tasks such as chatting, translation, summarization and batch text processing, Flash is recommended; for complex reasoning, long code, math problems and multi-step tasks, Pro is recommended, or the deep thinking mode should be enabled.
DeepSeek V4's specific parameter count, context window and billing standards are being continuously adjusted, and different sources give inconsistent figures, so it is recommended to consult the official documentation.
The ByteDance team published the paper 'Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression', pointing out that the DeepSeek V4 series models have phase sensitivity problems caused by chunked KV compression.
The Claude web interface has added an Agent mode and provides multiple model options including Claude, DeepSeek V4, Flash, and others.
Bilibili UP主 Xingyue organized a multi-dimensional AI model novel-writing evaluation covering six mainstream large models (Gemini 3.8 Flash High, Cloud Opus 5, GPT 5.6, DeepSeek V4, Qianwen 3.8, and GLM 5.3).
The demonstration uses the DeepSeek V4 model to generate science fiction novels; the model first displays its reasoning process and then outputs results node by node.
TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
The relay platform provides the DeepSeek V4 model, and selecting this model for calls consumes the relay platform's tokens.
Among Cline's free models is a mysterious model referred to as "Niulai" as well as DeepSeek V4.
DeepSeek V4 cost data is based on pricing before August 17, and prices may have increased after that date.
7 more timeline events
All Facts (20)
DeepSeek V4 adopts a proprietary hierarchical compressed attention strategy, featuring two alternating attention modes: CSA (Compressed Sink Attention) and HCA (Heavy Compressed Attention).
90%VerifiedDeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
80%VerifiedTileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
65%UnverifiedDeepSeek V4采用MUON优化器替代AdamW进行训练
85%UnverifiedDeepSeek V4引入了流形约束超链接（Manifold-constrained Hyperconnection，MHC）作为对传统残差连接的升级方案
85%UnverifiedDeepSeek V4支持百万级Token上下文窗口，并以MIT许可证发布
85%UnverifiedDeepSeek V4引入了MHC流形约束连接技术，用于保证深层网络训练稳定性
85%UnverifiedDeepSeek V4系列采用了MLA、MoE路由优化以及NSA等创新架构技术
85%UnverifiedDeepSeek V4采用了MoE（Mixture of Experts，混合专家）架构
75%UnverifiedThe DeepSeek team fine-tuned on DeepSeek v4 flash and developed their own visual encoder, DeepSeek ViT.
60%UnverifiedFor tasks such as chatting, translation, summarization and batch text processing, Flash is recommended; for complex reasoning, long code, math problems and multi-step tasks, Pro is recommended, or the deep thinking mode should be enabled.
50%UnverifiedDeepSeek V4's specific parameter count, context window and billing standards are being continuously adjusted, and different sources give inconsistent figures, so it is recommended to consult the official documentation.
50%UnverifiedThe ByteDance team published the paper 'Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression', pointing out that the DeepSeek V4 series models have phase sensitivity problems caused by chunked KV compression.
50%UnverifiedThe demonstration uses the DeepSeek V4 model to generate science fiction novels; the model first displays its reasoning process and then outputs results node by node.
50%UnverifiedThe relay platform provides the DeepSeek V4 model, and selecting this model for calls consumes the relay platform's tokens.
50%UnverifiedThe official deployment requirements for the DeepSeek V4 vision version demand 4 NVIDIA GB300 graphics cards to support full inference services.
50%UnverifiedThe Claude web interface has added an Agent mode and provides multiple model options including Claude, DeepSeek V4, Flash, and others.
50%UnverifiedBilibili UP主 Xingyue organized a multi-dimensional AI model novel-writing evaluation covering six mainstream large models (Gemini 3.8 Flash High, Cloud Opus 5, GPT 5.6, DeepSeek V4, Qianwen 3.8, and GLM 5.3).
50%UnverifiedDeepSeek released the first visual version weights of V4, with a model scale of 305B (approximately 305 billion parameters), along with a minimal inference implementation.
50%UnverifiedDeepSeek released the weights for the first vision version of V4, with a model scale of 305B (approximately 305 billion parameters).
50%