DeepSeek Previews New V4 Models That Rival World’s Top Ones in Performance

DeepSeek has just pulled the wraps off its latest AI model lineup, and it focuses on efficiency. The company has officially released DeepSeek-V4 Preview. Notably, it is open sourced with support for a massive 1M context length. The announcement came a day after OpenAI released GPT-5.5, which it says is its smartest model to date.

DeepSeek-V4 Preview goes live with focus on efficiency and scale

With the release of DeepSeek-V4 Preview, the company is looking to compete with both open and closed AI models around the globe. The new lineup includes DeepSeek-V4-Pro and DeepSeek-V4-Flash, each targeting different use cases while sharing the same core architecture.

DeepSeek-V4-Pro comes in with 1.6 trillion total parameters and 49 billion active ones. The company claims performance that rivals leading closed-source models, especially in reasoning-heavy tasks like coding, math, and STEM. It also reportedly leads most open models in world knowledge, trailing only top-tier systems like Gemini 3.1 Pro.

On the other hand, DeepSeek-V4-Flash focuses on speed and cost. With 284 billion total parameters and 13 billion actives, it is positioned as a faster and more affordable option. Interestingly, it still performs close to the Pro variant on simpler agent-based tasks, which could make it appealing for developers looking for efficiency without a big performance trade-off.

A push toward agentic AI and long-context workloads

DeepSeek is also focused on agentic AI, as it says the new models are designed to handle multi-step reasoning and integrate with tools like Claude Code, OpenClaw, and OpenCode. That suggests a clear shift toward real-world task execution rather than just text generation. If you are wondering about the pricing, take a look at the chart below:

Moving on, the company is also introducing a new attention mechanism with token-wise compression and sparse attention, which reportedly reduces compute and memory costs while handling long contexts more efficiently. The company has also shared a sample PDF generated by DeepSeek-V4-Pro, which you can see below:

With API access already live and open weights available, DeepSeek seems to be betting big on accessibility and scale.