DeepSeek was a viral sensation before OpenAI took over the conversation. Now, with a new release, it's looking to recapture that momentum.

On Friday, the Chinese lab DeepSeek rolled out previews of its new flagship models: DeepSeek-V4-Pro and DeepSeek-V4-Flash. DeepSeek released benchmarks that showcase how these models perform competitively against both open-source models and the top proprietary models from the frontier labs.

The new models offer performance upgrades from their predecessor, including:

• DeepSeek-V4-Pro:
  • 1.6 trillion total / 49 billion active parameters
  • Enhanced agentic capabilities and rich world knowledge that lead all open models, trailing only Gemini 3.1 Pro
  • Outperforms all current open models in reasoning tasks, including math, STEM, and coding, performing competitively with leading closed-source models, including GPT-5.4-High and Claude-Opus-4.6-Max
• DeepSeek-V4-Flash:
  • 284 billion total / 13 billion active parameters
  • Smaller parameter size for faster and cheaper performance

The biggest highlight is that, compared to DeepSeek-V3, the version of the model that went viral on its release in December 2024, because of its training efficiency and low cost, this model employs a different training architecture. That includes a new hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).

The result is better computational efficiency with reduced compute and memory costs despite remembering longer sequences, also known as expanding its context window. Specifically, it has a context length of one million tokens, addressing what the company describes as a key impediment in reasoning models, where computational constraints have traditionally made extended context lengths impractical.

A model's context window is a key determinant of its overall capabilities, as it governs how much past information the model can retain and reference at any given time. This is especially critical in the age of agentic workflows, where the ability to draw on a large volume of user information as context is essential. Despite these advancements, it achieves “significantly lower inference FLOPs,” a measure for computational costs, than DeepSeek-V3.2.

Interested users can try it on the DeepSeek website via the Expert Mode and Instant Mode options. To use it, you do have to either create an account or use the third-party sign-on options. It is also available in the API.

Our Deeper View

There is no denying the AI market's nearly insatiable hunger for everything bigger, faster, and more efficient. This appetite is only amplified by agentic technologies, which translate these capabilities into much more powerful results. Keeping costs at the center of their model is what made DeepSeek stand out the first time around, and given where things stand today, it is even more impressive that they have held firm to that philosophy and are doubling down on it. The DeepSeek-V4-Pro model costs $3.48 per million tokens, compared to $25 for Claude Opus 4.6 and $15 for GPT-5.4. And it achieves comparable benchmark scores. This release could put price pressure on the American labs, particularly as anxiety over competition from China continues to mount. That tension has reached the highest levels of government, with the White House recently accusing China of theft.