DeepSeek Releases Preview of New V4 AI Models - MLQ.ai

DeepSeek has released preview versions of its latest AI models: V4-Pro-Max with approximately 1.6 trillion parameters and V4 Flash with 284 billion parameters. Both are mixture-of-experts models and feature a 1 million token context window. The company claims these models rival or surpass OpenAI's GPT-5.2 and Google's Gemini 3.0 Pro on internal reasoning and coding benchmarks, though independent third-party evaluations have not yet been published.

The preview is available under usage caps for a limited period, according to the report. DeepSeek said the V4 models show improved reasoning capabilities and coding performance compared to its previous V-series releases. The earlier V3 model gained attention in early 2025 for showing competitive chart results at lower computational cost.

The release comes amid ongoing technical competition between Chinese and US AI labs, with racing capabilities often compared through performance parameters and reliance patterns. DeepSeek has not announced public window timing for maximum-resolution official large deployment including full testing framework access.