DeepSeek's AI Model: Bridging the Gap to Frontier Leaders

The Chinese AI laboratory DeepSeek has unveiled two preview versions of its latest large language model, DeepSeek V4. This highly anticipated release represents a significant advancement from last year’s V3.2 model and the acclaimed R1 reasoning model, which previously garnered considerable attention within the AI community.

According to the company, both DeepSeek V4 Flash and V4 Pro are architected as mixture-of-experts (MoE) models, each boasting an extensive context window of 1 million tokens. This capacity is sufficient to process substantial codebases or lengthy documents within a single prompt. The mixture-of-experts methodology enhances efficiency by activating only a subset of parameters for any given task, thereby reducing inference costs.

The Pro model features a formidable total of 1.6 trillion parameters, with 49 billion active, establishing it as the largest open-weight model currently available. This surpasses competitors such as Moonshot AI’s Kimi K 2.6 (1.1 trillion) and MiniMax’s M1 (456 billion), and more than doubles its predecessor, DeepSeek V3.2 (671 billion). The more compact V4 Flash model, by contrast, comprises 284 billion parameters, with 13 billion active.

DeepSeek states that both new V4 models demonstrate improved efficiency and performance compared to DeepSeek V3.2, a result of significant architectural enhancements. The company further claims that these models have nearly "closed the gap" with current leading models, both open-source and proprietary, across various reasoning benchmarks.

The company asserts that its V4-Pro-Max model outperforms its open-source counterparts on reasoning benchmarks and, on certain tasks, even surpasses OpenAI’s GPT-5.2 and Gemini 3.0 Pro. In competitive coding benchmarks, DeepSeek reports that both V4 models deliver performance "comparable to GPT-5.4."

However, the models appear to exhibit a slight deficit in knowledge tests when compared to cutting-edge frontier models, specifically OpenAI’s GPT-5.4 and Google’s latest Gemini 3.1 Pro. This discrepancy indicates a “developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months,” as acknowledged by the lab.

It is important to note that both V4 Flash and V4 Pro models currently support text-only processing. This contrasts with many of their closed-source competitors, which frequently offer multimodal capabilities, including the understanding and generation of audio, video, and images.

A distinctive advantage of DeepSeek V4 is its significantly more competitive pricing compared to existing frontier models. The V4 Flash model is priced at $0.14 per million input tokens and $0.28 per million output tokens, making it more affordable than GPT-5.4 Nano, Gemini 3.1 Flash, GPT-5.4 Mini, and Claude Haiku 4.5. The larger V4 Pro model is similarly competitive, costing $0.145 per million input tokens and $3.48 per million output tokens, undercutting Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4.

This launch occurs just one day after the U.S. government accused China of widespread industrial-scale intellectual property theft from American AI laboratories, allegedly utilizing thousands of proxy accounts. DeepSeek itself has faced accusations from Anthropic and OpenAI of "distilling," or effectively copying, their proprietary AI models.