DeepSeek V4 Unveiled: How Its Million-Token Context Redefines Open-Source LLMs

DeepSeek officially launched the preview of its V4 series, offering two variants: V4‑Pro with 1.6 trillion parameters, 49 billion activation‑layer neurons and 33 trillion tokens of pre‑training data; and V4‑Flash with 284 billion parameters, 13 billion activation‑layer neurons and 32 trillion tokens. Both models support a 1 million‑token context window and are available now via the official website and app.

Performance: Matching Top Closed‑Source Models

In agentic coding tests, V4‑Pro achieves the best results among open‑source models and is reported by internal users to outperform Anthropic’s Sonnet 4.5, approaching the quality of Opus 4.6 in non‑reasoning mode. In world‑knowledge benchmarks, V4‑Pro leads other open‑source models and trails only Gemini‑Pro‑3.1. On mathematics, STEM, and competition‑style coding tasks, V4‑Pro surpasses all publicly evaluated open models, matching the performance of leading proprietary systems.

Technical Breakthrough: Innovative Attention Mechanism

DeepSeek‑V4 introduces a new attention mechanism that compresses tokens along the token dimension and incorporates DeepSeek Sparse Attention (DSA). This design delivers world‑leading long‑context capability while substantially lowering computational and memory requirements compared with traditional approaches.

Economic Choice: V4‑Flash

For users seeking faster, more cost‑effective service, V4‑Flash offers a viable alternative. Although its world‑knowledge reservoir is slightly weaker than V4‑Pro, its reasoning ability is comparable, and it performs similarly on simple agent tasks.

API Service and Open‑Source Release

The DeepSeek API now supports both V4‑Pro and V4‑Flash, compatible with OpenAI ChatCompletions and Anthropic interfaces. The base URL remains unchanged; the model parameter should be set to deepseek-v4-pro or deepseek-v4-flash. Both variants support non‑reasoning and reasoning modes, with the reasoning_effort parameter controlling reasoning intensity (high/max). For complex agent scenarios, the reasoning mode with maximum intensity is recommended.

The model weights have been open‑sourced on Hugging Face and ModelScope, and the technical report is publicly available. The legacy API model names deepseek-chat and deepseek-reasoner will be deprecated on 2026‑07‑24, after which they will point to V4‑Flash’s non‑reasoning and reasoning modes respectively.