Chinese AI lab DeepSeek has unveiled previews of its latest large language model, DeepSeek V4, in two variants: Flash and Pro. This update to last year's V3.2 and the R1 reasoning model marks a significant leap forward, claiming to nearly "close the gap" with leading frontier models, both open-source and proprietary.
The new models, available on Hugging Face, are mixture-of-experts architectures with a context window of 1 million tokens—enough to handle entire codebases or large documents in a single prompt. By activating only a subset of parameters per task, MOE models achieve greater efficiency without sacrificing performance.
DeepSeek stated that architectural improvements enable both V4 Flash and Pro to outperform their predecessor V3.2 on reasoning benchmarks, offering a balance of speed and accuracy suitable for a wide range of applications. The Flash variant prioritizes throughput and lower latency, while Pro is optimized for maximum predictive quality.
The preview signals DeepSeek's continued ambition to compete with American and European AI labs. While the exact benchmark scores were not immediately disclosed, early indications suggest the models are closing in on state-of-the-art performance, a move that could intensify the global AI race.
Developers and researchers can access the preview models on Hugging Face to test their capabilities.