I thought I had to wait a few days, but today DeepSeek V4 finally arrived!! The preview version is officially launched and open sourced simultaneously. There are two versions in total: DeepSeeker V4 Pro: Benchmarking against top-level closed source models, 1.6T, 49B activation, context length 1M; DeepSeek-V4 Flash: A smaller and faster economy version, 284B, activated on the 13th, with a context length of 1M. Pro's performance is comparable to flagship closed source models, and its cost-effectiveness is simply amazing! I have a premonition that DeepSeek V4 will trigger a wave of "bookkeeping" this time! How much do you need to spend on similar performance? Compared to mainstream models: -GPT-5.4: Strong, but expensive -Claude Opus: Long texts are stronger, but more expensive -DeepSeek V4 Pro: Performance is close to the top tier, but the price is one order of magnitude lower. In the scenarios of Agent, RAG, and large-scale invocation, the cost advantage directly widens the gap. If you are making AI products, this is not an "optional" option, but a "mandatory" one. Low cost does not mean low capability. This is the first time that domestic models have truly entered the global main battlefield. Similar web development tasks were quickly completed.