DeepSeek‑V4 Launch: Open‑Source Model Matching Top Closed‑Source Performance with Dual Versions

On April 24, 2026 DeepSeek announced the preview release of DeepSeek‑V4, an open‑source large language model that supports up to 1 million tokens of context and offers both a high‑performance Pro version and a lightweight Flash version.

Core positioning : The model is marketed as affordable, efficient, strong in reasoning, and capable of long‑context processing, targeting developers, enterprises, and individual users with commercial‑ready, deployable, low‑cost solutions.

Version matrix :

DeepSeek‑V4‑Pro : 1.6 T total parameters, 49 B active parameters, trained on 33 T tokens, 1 M context length, positioned as “top‑tier performance, comparable to closed‑source ceilings”, expert mode.

DeepSeek‑V4‑Flash : 284 B total parameters, 13 B active parameters, trained on 32 T tokens, 1 M context length, positioned as “cost‑effective with strong inference and speed”, fast mode.

Key capabilities :

Inference performance : surpasses all publicly evaluated open‑source models on mathematics, STEM, and competitive coding benchmarks, approaching GPT‑5.4, Claude Opus 4.6, and Gemini‑3.1‑Pro; leads on MMLU‑Pro, SimpleQA‑Verified, Chinese‑SimpleQA, GPQA Diamond.

Agent ability : best open‑source agent performance, with “Agentic Coding” outperforming Sonnet 4.5 and nearing Opus 4.6 non‑thinking mode; toolchain fully compatible with Claude Code, OpenClaw, OpenCode, CodeBuddy; supports a “reasoning_effort” knob for high‑effort reasoning.

World knowledge : far ahead of other open‑source models, only slightly behind Gemini‑3.1‑Pro in knowledge accuracy, especially in Chinese comprehension and long‑text QA.

Long‑context handling : introduces a sparse attention mechanism called DSA that compresses tokens, dramatically reducing compute and memory while keeping efficiency leadership; enables one‑click processing of million‑word novels, papers, contracts, or codebases without segmentation.

API service : compatible with OpenAI ChatCompletions and Anthropic APIs; model names are deepseek‑v4‑pro and deepseek‑v4‑flash. Pricing per million tokens: Pro – input (cache hit) ¥1, input (miss) ¥12, output ¥24; Flash – input (hit) ¥0.2, input (miss) ¥1, output ¥2. Note that Pro’s throughput is limited by high‑end compute and will be reduced after the mid‑year launch of Ascend 950 super‑nodes.

Migration notice : the legacy endpoints deepseek‑chat and deepseek‑reasoner will be retired on July 24, 2026; they currently map to V4‑Flash non‑thinking and thinking modes.

Open‑source and deployment : weights are released on both HuggingFace and ModelScope; a full technical report discloses architecture, training methods, and evaluation data; deployment supports local, private, or cloud‑hosted environments and a wide range of hardware.

Target users : developers and R&D teams (codebase understanding, intelligent coding assistants, complex agents, low‑cost API); legal, finance, consulting (long contracts, reports, compliance QA); content and education (novel editing, literature review, course material synthesis); enterprise knowledge bases and workflow automation.

Immediate access : web UI at chat.deepseek.com, official mobile app, and direct API calls using the model names above.