HANGZHOU, April 24 (Xinhua) – Chinese AI company DeepSeek on Friday released and open sourced its long-awaited V4 model, which features superior performance in programming, world knowledge, and logical reasoning.
The technology startup, based in eastern China’s Hangzhou, says its new model’s Pro edition rivals the best open-source models in agent coding and has a significant lead in general knowledge behind the closed-source Gemini 3.1 Pro.
Additionally, the company announced that it ranks high on open source leaderboards in math, STEM, and competitive coding challenges.
Its Flash variant uses smaller parameter sizes and reduces activation overhead. Optimized for simple tasks and provides faster and more economical solutions.
Vals AI, a public LLM evaluation platform, states that DeepSeek V4 in X is “currently the #1 openweight model on the Vibe Code Benchmark, and it’s not even close.”
In a technical report published on Huggingface, DeepSeek said the new model validated the fine-grained scheme on both Nvidia GPU and Huawei Ascend NPU platforms.
Huawei announced on Friday that through close synergy with DeepSeek, its self-developed Ascend supernode products now support DeepSeek V4.
DeepSeek-V4 introduces a new attention mechanism featuring compression in the token dimension. By integrating this with DeepSeek Sparse retention, the company says the model supports context windows of over 1 million tokens, with significantly reduced compute and memory overhead compared to traditional approaches.
This week, Alibaba’s Qwen, Moonshot’s Kimi, and Tencent’s Hunyuan updated their own models.
Industry observers believe that from 2025 onwards, LLM iterations will enter an “ultra-short cycle,” with competition shifting from scale to real impact. Beyond the parameter count battle, the industry is focusing on inference efficiency, native multimodality, agent functionality, long context processing, and hallucination mitigation.
The goal, they say, is no longer just “being able to have a conversation” but “reliably completing complex tasks.” Meanwhile, the open source model is key to the developer ecosystem and global reach, driving innovation in coding and agents.
The Chinese tech company’s open-source large-scale AI models are ranked No. 1 in the world in terms of downloads, significantly lowering barriers to AI adoption, reducing usage costs, and increasing AI accessibility.
China’s average daily token calls soared from 100 billion in early 2024 to 140 trillion by March 2026, according to data from the National Data Bureau. ■
