DeepSeek

DeepSeek V4.1 Flash is DeepSeek's open-weights, ultra-efficient vision-language Mixture-of-Experts (MoE) foundation model. Built on a novel Causal Encoder-Decoder (CED) architecture, it distributes 552B total parameters while activating only 8B parameters during prefill and 16B parameters during decode. Featuring native visual-text joint embedding, Compressed Sparse Attention, Engram memory, and a 4x KV-cache memory reduction, it effortlessly processes up to a 1M-token context window at high throughput. Engineered for extreme cost efficiency and frontier agentic coding, DeepSeek V4.1 Flash powers autonomous software engineering, multimodal document RAG, UI browser automation, and high-concurrency API integrations.

DeepSeek-V4-Pro-0813 is a flagship MoE model released on August 13, 2026, serving as the official GA release built for complex logical reasoning, code construction, and AI Agent collaboration. Architecture & Performance: 1.6T parameters (49B active), built-in DSpark speculative decoding, with significantly enhanced inference throughput. Compatible with OpenAI/Anthropic APIs, deeply integrated with DeepSeek Harness, and delivers outstanding performance on Terminal Bench.

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts (MoE) model from DeepSeek, featuring 284B total parameters with 13B activated per token and a 1-million-token context window. Designed for fast inference and high-throughput workloads, it delivers strong reasoning and coding capabilities while maintaining excellent cost efficiency. Its hybrid attention architecture enables efficient long-context processing. The model supports **high** and **xhigh** reasoning levels, with **xhigh** representing the maximum reasoning effort. It is ideal for coding assistants, conversational systems, and agentic workflows where responsiveness, scalability, and cost efficiency are essential.