open source ai disrupts monopoly
Up next
Author
Tags
Share article
The post has been shared by 0 people.
Facebook 0
X (Twitter) 0
Pinterest 0
Mail 0

China’s DeepSeek has released V4, a powerful open-source AI model that’s shaking up the industry. For years, big tech companies like Google, Microsoft, and Anthropic have dominated the AI space. DeepSeek’s latest model is challenging that control directly.

DeepSeek V4 comes in two versions. The Pro version has 1.6 trillion total parameters but only uses 49 billion when running. The Flash version is smaller, with 284 billion total parameters and 13 billion active ones. Both versions support a one-million-token context window, meaning they can process enormous amounts of text at once.

DeepSeek V4 packs 1.6 trillion parameters and a one-million-token context window into two powerful versions.

The model’s benchmark scores are turning heads. V4-Pro scored 91.0 on MMLU-Pro, a tough knowledge test. It also scored 64.5 on the MATH benchmark, leading all open-source models. On SWE Verified, a coding test, it resolved 80.6% of problems. That puts it close to top closed models like Claude Opus 4.6.

Cost is where DeepSeek’s V4 really stands out. V4-Pro costs $1.74 per million input tokens. That’s 10 times cheaper than Claude Sonnet 4.6 and 17 times cheaper than Opus 4.7. Compared to GPT-5.5, users could save around 85%. Those savings are significant for developers and businesses running large-scale AI tasks.

DeepSeek also made serious efficiency improvements. V4 uses only 10% of the KV cache that its predecessor V3.2 needed in one-million-token scenarios. Its inference cost stays flat as token count grows, unlike older models that got more expensive. A new system called Deeply Sparse Attention helps the model handle long texts without a computing explosion.

The model performs well in real-world tasks. It’s been tested in app building, coding agents, and even generating a Minecraft clone. It’s already integrated with tools like Code, Claw, and OpenCode. Developers can also adjust the model’s thinking intensity using a reasoning_effort parameter set to low, high, or max levels.

However, testers noted it still falls short in creative writing and polished outputs. Retailers and businesses exploring AI tools have increasingly turned to AI-powered security systems to address the over $13 billion lost annually to shoplifting, highlighting how cost-effective AI deployment is becoming critical across industries.

DeepSeek released V4 fully on Hugging Face with a complete system card. Both the Pro and Flash versions are available as open-weight models released under the MIT License, allowing developers to freely modify and deploy them. It’s still trailing GPT-5.5 and Claude Opus 4.7 on the hardest benchmarks by a few months. But its combination of performance and low cost is making waves across the AI industry.

References

You May Also Like
reinforcement learning in ai

Reinforcement Learning: The Bold Framework Behind AI’s Most Jaw-Dropping Achievements

AI learns like toddlers touching hot stoves – except it’s beating world champions and making million-dollar trades while humans watch nervously.
entropy rich neural networks advantage

Why Entropy-Rich Neural Networks Dramatically Outperform Traditional Training Methods

Entropy-rich neural networks crush traditional training methods in speed and accuracy—but the real surprise is how they reshape memory capacity entirely.
cognitive processing speed disparity

Your Brain’s Bottleneck: Why We Think 100 Million Times Slower Than We See

Your brain absorbs a billion bits per second but thinks at just 10. The bizarre bottleneck behind this 100-million-fold gap changes everything you assume.