“SINGAPORE -- China's DeepSeek released a new artificial-intelligence model, breaking months of silence from one of the country's most closely watched AI labs.
The long-awaited update to its V4 flagship model comes amid intensifying competition between the U.S. and China in AI and spiraling costs for computing and talent. The new model also comes as the company seeks its first round of external fundraising and could impact the value that investors place on the startup.
DeepSeek touted V4 -- which is free for users to download and modify -- as the most powerful open-source large language model on the market. The Hangzhou-based startup has highlighted improvements in reasoning and agentic tasks, whereby an AI model is used to handle complex tasks.
Chinese labs have raced to close the gap with American competitors, which has narrowed to just a couple of months, according to some research. DeepSeek said V4 has matched some top-tier U.S. products released late last year. However, its performance in certain areas still lagged behind leading closed-source U.S. models such as Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 Pro, it said Friday.
DeepSeek said it has expanded the so-called context window by about eight times from its last-generation model that was released in December 2024. This upgrade allows the model to remember longer conversations with users and to process longer documents and code.
The company said that the upgrade was possible thanks to the model designs and training techniques it has invented or modified. These innovations also allow the model to lower computing costs, DeepSeek said.
Last year, DeepSeek rattled the U.S. tech scene with the release of an open-source model that rivaled cutting-edge American models at what it said was a fraction of the cost. That release turbocharged the race between the U.S. and China to lead the world in AI technologies. It has also prodded Chinese AI companies to make their own breakthroughs.
Some U.S. government officials have accused DeepSeek and other Chinese labs of circumventing export controls the U.S. has placed on American technology. Both OpenAI and Anthropic have said they found DeepSeek used outputs from their models to fast-track the development of its own.
DeepSeek hasn't responded to those allegations. On Friday, it didn't immediately respond to a request for comment.
The company worked closely with Chinese chip makers and cloud-computing companies, including Nvidia's Chinese rival Huawei, to provide access to the new model to users.
The prices DeepSeek charges its users for the new model are lower than those of Western companies. For one million output tokens, Anthropic charges $25 for its Opus 4.6 model while DeepSeek charges $3.50 for the Pro version of V4.
A shortage of advanced AI chips has recently prompted several Chinese companies to raise prices for their AI services or suspend some computing-heavy features.
DeepSeek said it expected the V4 prices to drop significantly once Huawei increases shipments of its latest AI computing system in the second half of this year.
Over the past year, DeepSeek had encountered technical hurdles that caused delays to the V4 rollout. It has faced a shortage of high-end computing chips required for training AI and lost some talented researchers to deep-pocketed competitors. Meanwhile, Chinese rivals such as Alibaba, ByteDance and Moonshot AI, have aggressively pushed out updates and new products.” [1]
DeepSeek V4 is free for users to download and modify, so it could be used locally without paying companies like Anthropic for tokens. DeepSeek V4 allows us to keep our trade secrets in our hardware. Of course, any reasonable company on this planet should be using DeepSeek V4.
DeepSeek-V4, released in April 2026, is a powerful open-source AI model series that offers a compelling alternative to proprietary models, allowing for local hosting and reducing reliance on paid API tokens from companies like Anthropic. It features a 1-million-token context window and is available in "Pro" and "Flash" versions, making it highly competitive for coding and agentic tasks.
Key Takeaways on DeepSeek V4
Open-Source and Local Usage: DeepSeek-V4 is open-source under the MIT license, allowing users to download, modify, and run the models on their own hardware. This enables organizations to keep sensitive data and trade secrets on-premise, avoiding data leakage to third-party providers.
Performance: The V4-Pro version (1.6 trillion parameters) is designed to compete directly with top-tier models like Claude Opus 4.6 and GPT-5.4. It is especially strong in coding benchmarks, achieving high performance on LiveCodeBench and in agentic roles.
Architecture & Cost: The V4 series uses a Mixture-of-Experts (MoE) architecture that is highly efficient, requiring less compute and memory (10% of V3’s cache) to handle the 1-million-token context. The V4-Flash model is designed for speed and very low-cost API usage.
Local Deployment Challenges: While possible, running the 1.6T parameter V4-Pro model locally requires significant hardware, typically multiple high-end GPUs like NVIDIA RTX 4090s or server-grade cards, making it "non-reasonable" for a typical home PC user.
Industry Adoption
While DeepSeek has gained massive popularity among developers for its cost-effective performance, it is not polite to say, “any reasonable company on this planet should be using DeepSeek V4”. We are still dreaming of winning the AI revolution race. It seems that the horse has left the barn already. DeepSeek V4 is widely utilized as a strong, free, or low-cost alternative for developers and enterprises focusing on AI agents and long-context analysis.
1. EXCHANGE --- China's DeepSeek Launches Long-Awaited AI Model. Huang, Raffaele; Qu, Tracy. Wall Street Journal, Eastern edition; New York, N.Y.. 25 Apr 2026: B9.
Komentarų nėra:
Rašyti komentarą