HANGZHOU, April 24 (Xinhua) — Chinese AI firm DeepSeek on Friday released and open-sourced its highly anticipated V4 model, which features good performance in programming, world knowledge and logical reasoning.
The new model’s Pro edition matches the best open-source models in agentic coding and significantly leads in general knowledge, second only to the closed-source Gemini 3.1 Pro, according to the tech startup based in Hangzhou in east China.
Moreover, it ranks among the top positions in open-source leaderboards in math, STEM and competitive coding challenges, the company announced.
Its Flash variant utilizes a smaller parameter size and reduces activation overhead. Optimized for straightforward tasks, it offers a more rapid and economical solution.
Vals AI, a public LLM evaluation platform, noted on X that DeepSeek V4 is “now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close.”
In a technical report released on Huggingface, DeepSeek said that the new model has validated a fine-grained scheme on both Nvidia GPUs and Huawei Ascend NPUs platforms.
Huawei announced Friday that homegrown Ascend super-node products now support DeepSeek V4, after close synergy with DeepSeek.
DeepSeek-V4 introduces a new attention mechanism featuring compression in the token dimension. By integrating this with DeepSeek Sparse Attention, the model supports a context window of over 1 million tokens, drastically lowering compute and memory overhead relative to conventional approaches, the firm said.
This week, Alibaba’s Qwen, Moonshot’s Kimi and Tencent’s Hunyuan updated their own models.
Industry insiders believe that since 2025, LLM iteration has entered an “ultra-short cycle,” with competition shifting from scale to real-world effectiveness. The industry has moved beyond parameter-count wars to focus on inference efficiency, native multimodality, agent capabilities, long-context processing and hallucination mitigation.
They said the goal is no longer just “being able to converse” but “reliably completing complex tasks.” Meanwhile, open-source models have become key to developer ecosystems and global expansion, driving innovation in coding and agents.
Chinese tech firms’ open-source large AI models rank first globally in downloads, significantly lowering barriers to AI adoption, reducing usage costs and enhancing AI accessibility.
Data from the National Data Administration shows China’s average daily token calls had surged from 100 billion in early 2024 to 140 trillion by March 2026. ■
