DeepSeek is back: China's AI claims to surpass ChatGPT and Gemini in key benchmarks

Chinese AI startup DeepSeek has officially released preview versions of its highly anticipated DeepSeek-V4 models. The much awaited update from DeepSeek comes more than a year after its R1 and V3 models went viral last year and broke all notions of US supremacy in the AI race.

The latest model from DeepSeek comes with significant architectural upgrades, multiple reasoning modes, and a massive one-million token context window.

Both models support an ultra-long context length of one million tokens (approximately 750,000 words)

The new DeepSeek-V4 models come in three reasoning modes: Non-think, Think High and Think Max. DeepSeek says the Non-think mode is aimed for daily tasks and low-risk decisions while Think High is for questions that require Complex problem-solving and planning. Meanwhile, the Think Max is for handling hardest coding and math problems.

In a Hugging Face page for the model, DeepSeek says that the V-4 Pro Max and V4 Pro “significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks”

DeepSeek-V4-Pro-Max leads in coding and mathematical performance, topping the Apex Shortlist, a benchmark focused on high-difficulty reasoning and problem-solving, with a score of 90.2%. It also achieves a Codeforces rating of 3206, which shows a strong real-world competitive programming ability, and ties for first place on SWE Verified, a benchmark which evaluates performance on practical software engineering tasks.

However, the model lags behind its American counterparts in general knowledge and broader reasoning. Gemini 3.1 Pro leads on SimpleQA-Verified, a benchmark designed to test factual accuracy and question answering, while GPT-5.4 ranks highest on Terminal Bench 2.0, which measures how effectively models can use tools and operate in agent-like environments.

DeepSeek says the V4-Pro-Max achieves these results while being far more efficient, using nearly 10 times less memory than its V3.2 model when handling long inputs.