🐳 DeepSeek-V4: The Long-Awaited Release from the Chinese Startup

DeepSeek has officially launched the new version of its AI model, V4, in two variants: Pro and Flash, both with a default context window of 1 million tokens.

✅ The Pro model scores on par with GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6 in benchmarks. It's especially strong in general knowledge, STEM, and autonomous coding.

✅ The Flash version isn't far behind the flagship, while being faster and cheaper.

✈️ Regular users will feel the improvements right away in any task: DeepSeek matches the top Western AI in both speed and intelligence.

💡 Under the hood, the model packs plenty of architectural improvements that make V4 far more efficient: it uses just 27% of the computing power compared to the previous V3.2.

Context is tightly compressed thanks to an optimized attention mechanism—the model keeps only the essential information and can see 10x more data with the same resources.

💻 DeepSeek-V4 is optimized for Chinese Huawei chips—that process was one of the reasons the release took a while. Now, though, the company is carefully but almost openly talking about independence from Nvidia hardware and will keep fine‑tuning for local accelerators, which will make V4 even faster and cheaper down the line.

➡️ You can already try it in DeepSeek Chat for free.

Do you use DeepSeek?

❤️ — Yes, all the time
🤔 — No, it doesn't suit me
🔥 — No, but now I will!

@hiaimediaen