DeepSeek Releases New V4 Model: Million-Token Ultra-Long Context Becomes Standard
Chinese AI startup DeepSeek announced the release of its new series model, DeepSeek-V4, with a preview version now open-sourced. The V4 models boast ultra-long context capabilities of one million tokens, establishing a leading position in agent capabilities, world knowledge, and reasoning performance within China and the open-source domain.
📋 Article Processing Timeline
- 📰 Published: April 24, 2026 at 17:41
- 🔍 Collected: April 24, 2026 at 18:02 (21 min after Published)
- 🤖 AI Analyzed: April 24, 2026 at 21:14 (3h 12m after Collected)
Central News Agency
(Central News Agency, Taipei 24th) Chinese artificial intelligence (AI) startup DeepSeek announced today on its WeChat official account that the preview version of its new series model, DeepSeek-V4, has been officially launched and open-sourced simultaneously. DeepSeek claims that V4 possesses ultra-long context of one million tokens, achieving a leading position in agent capabilities, world knowledge, and reasoning performance both domestically and in the open-source domain.
DeepSeek officials stated that the V4 model is divided into two versions, Pro and Flash, with DeepSeek-V4-Flash being the faster and more efficient economic choice.
DeepSeek officials pointed out that V4 pioneered a new attention mechanism that compresses in the token dimension, combined with DSA Sparse Attention (DeepSeek Sparse Attention), to achieve globally leading long-context capabilities. It also significantly reduces computational and memory requirements compared to traditional methods. "From now on, 1M (one million) context will be the standard for all official DeepSeek services."
DeepSeek officials also stated that DeepSeek-V4-Pro significantly outperforms other open-source models in world knowledge assessment, falling only slightly behind the top closed-source model, Gemini-Pro-3.1.
This marks the launch of the V4 model by DeepSeek more than a year after the release of its V3 model at the end of 2024.
Huawei's WeChat official account stated on the 24th that the Ascend supernode, based on the Ascend 950 AI chip, will fully support DeepSeek's V4 version.
The day before DeepSeek-V4's preview release, the U.S. government accused China in a memorandum of industrial-scale theft of intellectual property from U.S. AI labs.
Reuters quoted Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), as writing in the memo: "The U.S. government has information indicating that foreign entities, primarily located in China, are deliberately engaging in industrial-scale operations to distil cutting-edge U.S. AI systems."
"Distillation" refers to using output data from larger AI models to train smaller AI models, a method that helps reduce costs when training powerful new AI tools.
In February this year, U.S. AI company Anthropic stated that DeepSeek, Moonshot AI, and MiniMax illegally extracted technical capabilities from its chatbot Claude, directly accusing them of industrial-scale intellectual property theft. (Edited by Chen Kai-yu / Yang Sheng-ju) 1150424
Stand with the facts, your sponsorship is the power to protect press freedom.
Download the Central News Agency's "First-hand News" APP to stay updated with the latest news.
The text, images, and videos on this website may not be reproduced, publicly broadcast, or publicly transmitted and used without authorization.
(Central News Agency, Taipei 24th) Chinese artificial intelligence (AI) startup DeepSeek announced today on its WeChat official account that the preview version of its new series model, DeepSeek-V4, has been officially launched and open-sourced simultaneously. DeepSeek claims that V4 possesses ultra-long context of one million tokens, achieving a leading position in agent capabilities, world knowledge, and reasoning performance both domestically and in the open-source domain.
DeepSeek officials stated that the V4 model is divided into two versions, Pro and Flash, with DeepSeek-V4-Flash being the faster and more efficient economic choice.
DeepSeek officials pointed out that V4 pioneered a new attention mechanism that compresses in the token dimension, combined with DSA Sparse Attention (DeepSeek Sparse Attention), to achieve globally leading long-context capabilities. It also significantly reduces computational and memory requirements compared to traditional methods. "From now on, 1M (one million) context will be the standard for all official DeepSeek services."
DeepSeek officials also stated that DeepSeek-V4-Pro significantly outperforms other open-source models in world knowledge assessment, falling only slightly behind the top closed-source model, Gemini-Pro-3.1.
This marks the launch of the V4 model by DeepSeek more than a year after the release of its V3 model at the end of 2024.
Huawei's WeChat official account stated on the 24th that the Ascend supernode, based on the Ascend 950 AI chip, will fully support DeepSeek's V4 version.
The day before DeepSeek-V4's preview release, the U.S. government accused China in a memorandum of industrial-scale theft of intellectual property from U.S. AI labs.
Reuters quoted Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), as writing in the memo: "The U.S. government has information indicating that foreign entities, primarily located in China, are deliberately engaging in industrial-scale operations to distil cutting-edge U.S. AI systems."
"Distillation" refers to using output data from larger AI models to train smaller AI models, a method that helps reduce costs when training powerful new AI tools.
In February this year, U.S. AI company Anthropic stated that DeepSeek, Moonshot AI, and MiniMax illegally extracted technical capabilities from its chatbot Claude, directly accusing them of industrial-scale intellectual property theft. (Edited by Chen Kai-yu / Yang Sheng-ju) 1150424
Stand with the facts, your sponsorship is the power to protect press freedom.
Download the Central News Agency's "First-hand News" APP to stay updated with the latest news.
The text, images, and videos on this website may not be reproduced, publicly broadcast, or publicly transmitted and used without authorization.