DeepSeek-V4: The Future of Open-Source AI Unveiled
Chinese AI lab DeepSeek released preview versions of its long-awaited V4 model on Wednesday, open-sourcing two variants that claim near-parity with the world's most powerful closed-source AI systems at a fraction of the cost. The launch marks DeepSeek's most ambitious release since its R1 model rattled global AI markets in early 2025.
Two Models, One Million Tokens
DeepSeek announced the release on its official X account, writing that "DeepSeek-V4 Preview is officially live & open-sourced" and declaring the arrival of "the era of cost-effective 1M context length." The company published model weights and a technical report on Hugging Face the same day.
The flagship DeepSeek-V4-Pro packs 1.6 trillion total parameters with 49 billion activated per token, while the lighter DeepSeek-V4-Flash comes in at 284 billion total parameters with 13 billion activated. Both use a Mixture-of-Experts architecture and support a context window of one million tokens — among the largest in the industry. A new hybrid attention mechanism combining Compressed Sparse Attention and Heavily Compressed Attention allows V4-Pro to use just 27% of the inference compute and 10% of the memory cache required by its predecessor, DeepSeek-V3.2, at the same context length.
Benchmarks and Pricing
According to DeepSeek's own evaluations, V4-Pro in its maximum reasoning mode scored 93.5 on LiveCodeBench and achieved a 3,206 Codeforces rating, topping Google's Gemini-3.1-Pro and OpenAI's GPT-5.4 on both measures. On the widely watched SWE-bench Verified, V4-Pro reached 80.6% — within a percentage point of Anthropic's Opus-4.6, which leads at 80.8%. The technical report describes V4-Pro-Max as "the best open-source model available today," though it trails closed-source leaders on several knowledge and agentic benchmarks.
Pricing may prove as disruptive as the benchmarks. V4-Flash is listed at $0.14 per million input tokens and $0.28 per million output tokens, while V4-Pro costs $1.74 and $3.48 respectively, according to the company's published rate card. That makes V4-Pro roughly three times cheaper than GPT-5.5 on input and nearly nine times cheaper on output, according to Apidog's analysis.
A Year After the R1 Shock
The release caps a turbulent development cycle. DeepSeek originally targeted a February 2026 launch but delayed twice, in part due to challenges with Huawei Ascend chip infrastructure, Reuters reported in April, citing The Information. The model's deep compatibility with domestic Chinese chips represents what one report called "a critical step forward for China's AI industry on the path to 'de-CUDA-ization.'"
Major Chinese cloud providers including Alibaba, ByteDance, and Tencent have reportedly pre-ordered hundreds of thousands of next-generation AI chips in anticipation of offering V4 through their platforms. The models are available immediately through DeepSeek's API and chat interface, with the MIT-licensed weights free for download.
