DeepSeek V4: Months Behind Giants, Yet Boasting Intergalactic Speed

by

The cosmos of technology keeps turning like a very expensive washing machine, and China’s DeepSeek is waving its towel in the air and claiming that its shiny new V4 models are only months away from catching up with the giant-sized OpenAI and the slightly more sensible Google. If you squint hard enough, you can almost hear the universe sigh and mutter, “Well, perhaps we should all calm down and let the robots finish their tea.”

Summary

  • DeepSeek rolled out V4-Pro and V4-Flash preview models, insisting they’re a mere three to six months behind the reigning titans OpenAI and Google.
  • The open-source V4-Pro leads rival open models in maths and coding benchmarks, while V4-Flash promises similar reasoning with the speed of a caffeinated comet and a bill that won’t bankrupt your budget of aspirations.
  • The rollout follows the impact of DeepSeek-R1 and arrives amid regulatory risings and a narrowing US-China AI performance gap, according to the Stanford AI Index 2026 report.

Around a year after its last splash of bravado, the Hangzhou-based startup unveiled the DeepSeek V4 Pro and V4 Flash preview models on a Friday that evidently decided to think it over and then announced a leap anyway, because nothing screams progress like pretending you’ve already landed on the Moon while still in the launch queue.

Performance narrows gap with closed models

The DeepSeek V4 Pro and V4 Flash models parade themselves as top-tier contenders, like very polite but ambitious ducks. DeepSeek states that V4 Pro leads all open-source models in maths and coding benchmarks, while acknowledging that it lags behind closed systems like Google’s Gemini 3.1 Pro in general knowledge-though the gap, of course, is embarrassingly small, unless you’re measuring by which planet you’re on.

DeepSeek estimates they are now only three to six months behind leading models. The V4 Flash model is designed for speed and efficiency, offering similar reasoning to the Pro but at a lower cost for large-scale adventures, or at least for large-scale spreadsheets that never sleep.

This release follows DeepSeek R1, which some, including the rather perceptive Marc Andreessen, considered a turning point in AI. That release showed high-level reasoning could be achieved with less capital, as DeepSeek claimed a training cost of under $6 million. The architecture radiates efficiency, though a few analysts are skeptical of that tiny number-much like the number on a receipt that suddenly becomes mystical after a few billable hours.

The rapid ascent of DeepSeek has invited scrutiny. Because AI is now a kind of global sport with a lot of money at stake, these models are under heavy review and careful gazing by people who like to wear serious glasses.

Some regions, including Taiwan, Australia, and parts of the U.S., have restricted the use of earlier DeepSeek models due to data privacy and national security concerns. In the Great Hall of Policy, the broom closet is apparently where most decisions sleep.

The Stanford AI Index 2026 report confirms that while the U.S. still leads in high-impact patents and model breakthroughs, China has closed the gap in publication volume and industrial applications-though the coffee is still the same price everywhere and the potential for misprints remains universal.

Competition intensifies across open and closed AI models

DeepSeek continues to champion an open-source approach, inviting developers to tinker, tweak, and possibly break things in glorious, educational ways. This places them in competition with Google’s recently released Gemma 4, which concentrates on agent-style workflows and task automation-because who doesn’t want a robot therapist that files your emails while you pretend to work?

As OpenAI hones its closed, enterprise-grade systems, V4 suggests that the choice between open-source accessibility and closed-door performance is becoming a deliciously thorny decision for developers who enjoy choosing their own adventure but dislike paying for the map.

Read More

2026-04-24 14:18