DeepSeek V4 Pro

אודות DeepSeek AI

הוקם בשנת 2023 · Hangzhou, China

DeepSeek AI is a Chinese AI research lab founded in 2023 by Liang Wenfeng, the founder of the High-Flyer quantitative hedge fund. The lab is funded primarily by High-Flyer's profits and operates independently of major Chinese tech conglomerates. DeepSeek's mission is to build open frontier AI: every flagship model has been released with open weights and a permissive license. Major releases include DeepSeek LLM (late 2023), DeepSeek-V2 (May 2024, MoE), DeepSeek-V3 (December 2024, 671B MoE / 37B active), DeepSeek-R1 (January 2026 family, reasoning), DeepSeek V3.1 (early 2026) and the DeepSeek V4 family released as preview April 24, 2026. DeepSeek's models repeatedly top open-weights leaderboards on coding, math and reasoning at a fraction of the training cost claimed by Western labs, and the team is credited with popularising large-scale Reinforcement Learning from Verifiable Rewards.

ארכיטקטורה

Sparse Mixture-of-Experts Transformer (frontier open-weights)

DeepSeek-V4-Pro was released April 24, 2026 as the flagship of the V4 family. It is a Sparse MoE Transformer with 1.6T total parameters and 49B activated per token, supporting a native 1M-token context window with up to 384K-token max output. The model was trained on the lab's expanded GPU cluster using DeepSeek's signature recipe: large-scale pretraining on a multi-trillion-token mixture of web text, code, books, scientific papers and curated math/STEM data, followed by extensive Reinforcement Learning from Verifiable Rewards (RLVR) on math, coding and tool-use trajectories. Architectural innovations introduced in V3 - Multi-head Latent Attention (MLA), DeepSeekMoE with fine-grained expert specialization and shared experts, and FP8 mixed-precision training - are retained and refined. V4 Pro is published with open weights under a permissive license and runs natively in both server and inference frameworks such as vLLM and SGLang. DeepSeek treats the V4 launch as a preview phase and has announced that the older deepseek-chat and deepseek-reasoner endpoints will be deprecated on July 24, 2026.

פרמטרים

1.6T total / 49B active per token

הקשר

1,048,576 טוקנים

יכולות

• 1M token native context window with 384K max output
• ~81% SWE-bench Verified - rivals top closed-source models
• Top open-weights scores on Math/STEM/Coding benchmarks
• 1.6T MoE / 49B active parameters
• FP8-trained for compute efficiency
• Multi-head Latent Attention for memory-efficient long context
• Function calling and structured JSON output
• Cache hit pricing at $0.0145 per million tokens enables cheap multi-turn agents
• Available via DeepSeek API, OpenRouter, Together, Fireworks and self-hosted with vLLM/SGLang
• Open weights under a permissive license
• Best for: open-weights coding agents, long-document analysis, cost-efficient reasoning workloads, on-premise enterprise deployments.

הדרכה ורישיון

Pretrained on a multi-trillion-token mixture of web text, code, books, scientific papers and curated math/STEM data. Post-training applies large-scale Reinforcement Learning from Verifiable Rewards (RLVR) on math, coding and tool-use tasks, plus supervised fine-tuning and instruction tuning. Knowledge cutoff approximately early 2026.

רישיון: Open weights under a permissive license that allows commercial use. Hosted API access via deepseek.com.

בדיקות בטיחות: DeepSeek publishes model cards but provides limited external red-teaming. Safety filters are lighter than Western frontier labs; deployers are responsible for downstream alignment.