DeepSeek unveils new AI model aiming to narrow the gap with top frontier systems

Chinese AI research lab DeepSeek has released two preview versions of its latest large language model, DeepSeek V4, marking a significant update to last year’s V3.2 model and the accompanying R1 reasoning system that drew widespread attention across the AI industry.

According to the company, both DeepSeek V4 Flash and V4 Pro are built using a mixture-of-experts architecture and feature context windows of up to 1 million tokens. This extended capacity allows users to input large datasets such as extensive codebases or lengthy documents. The mixture-of-experts design works by activating only a subset of parameters for each task, helping reduce inference costs while maintaining performance.

The V4 Pro model contains 1.6 trillion parameters, with 49 billion active at any given time, making it the largest open-weight model currently available. This places it ahead of competing systems such as Moonshot AI’s Kimi K 2.6, which has 1.1 trillion parameters, MiniMax’s M1 at 456 billion parameters, and more than doubles the size of DeepSeek V3.2, which stood at 671 billion parameters. The smaller V4 Flash model includes 284 billion parameters, of which 13 billion are active.

DeepSeek reports that both models deliver improved efficiency and performance over V3.2, attributing the gains to architectural enhancements. The company states that these upgrades have nearly eliminated the performance gap between its models and today’s leading AI systems, across both open-source and proprietary categories, particularly in reasoning benchmarks.

The lab further claims that its V4-Pro-Max configuration outperforms other open-source competitors on reasoning tasks and even exceeds models like GPT-5.2 and Gemini 3.0 Pro on certain benchmarks. In programming-focused evaluations, DeepSeek indicates that both V4 variants perform at levels comparable to GPT-5.4.

Despite these advances, the models appear to trail slightly behind top-tier systems in knowledge-based assessments, particularly when compared with GPT-5.4 and Gemini 3.1 Pro. DeepSeek noted that this suggests a development curve that lags behind cutting-edge frontier models by roughly three to six months.

Both V4 Flash and V4 Pro are limited to text-based capabilities, unlike several closed-source models that support multimodal inputs and outputs, including audio, video, and image processing.

A notable aspect of the release is pricing. DeepSeek positions V4 as significantly more cost-effective than many current frontier systems. The V4 Flash model is priced at $0.14 per million input tokens and $0.28 per million output tokens, undercutting offerings such as GPT-5.4 Nano, Gemini 3.1 Flash, GPT-5.4 Mini, and Claude Haiku 4.5. Meanwhile, the V4 Pro model costs $0.145 per million input tokens and $3.48 per million output tokens, and it also comes in below competitors such as Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4.

The release follows closely on the heels of renewed geopolitical tensions in the AI sector, after the United States accused China of conducting large-scale intellectual property theft targeting American AI labs using thousands of proxy accounts. DeepSeek itself has previously faced allegations from Anthropic and OpenAI that it engaged in “distillation,” a process described as effectively copying the behaviour of other AI models.