Skip to main content

DeepSeek V4 Final Version Leaked, May Launch Next Monday to Take on Kimi K3

Just when everyone thought the AI spotlight was fixed on Kimi K3, DeepSeek decided to crash the party. Test videos and real-machine demos of the final version of DeepSeek V4 have quietly surfaced on social platforms, and multiple sources suggest the model could drop as early as next Monday—the first Monday of mid-July. The timing is no coincidence: right on the heels of K3's hype, DeepSeek is aiming to steal the show.

According to leaked test data and feedback from overseas bloggers, the official V4 is turning heads. Preliminary results put it on par with Claude Opus 4.8, especially in 3D generation, where it reportedly leaves Opus 4.8 in the dust. Agent behavior and coding stability have also seen major upgrades. Some bold voices even claim that in specific tests—like game code generation—V4 has already surpassed Fable5 and GPT-5.6. One blogger demonstrated a "Squid Game" mini-game coding test: V4 churned out 3,000 lines of code at a total cost of just $0.12. That's not just impressive; it's a cost-effectiveness bombshell.

But it's not all about raw power. V4 has also refined its interaction style. Testers noticed that its chain-of-thought (CoT) now feels more like Claude's—no more endless walls of code. Instead, it presents summary steps like "writing core functions" and "writing main scene code," making it far easier to follow. Some analysts think V4 may have borrowed a page from Fable's playbook, resulting in cleaner, more concise outputs.

Image

Looking back at DeepSeek's track record, they've always played the "work quietly and undercut everyone" card. If V4 can deliver top-tier performance at a price an order of magnitude lower, we might be looking at "DeepSeek Moment 2.0"—a sequel to the V3 launch that shook the industry. The company had previously announced a mid-July release, so next Monday fits the timeline. Of course, we'll have to wait for official confirmation.

Key Points

  • DeepSeek V4 final version test videos leaked online; official release possibly next Monday.
  • Benchmarks show V4 matching Claude Opus 4.8, with superior 3D generation and coding stability.
  • Cost efficiency highlighted: 3,000 lines of game code generated for $0.12.
  • Chain-of-thought style improved for better readability, similar to Claude.
  • Could trigger a new "DeepSeek Moment" if pricing remains aggressive.