DeepSeek releases V4 Flash Vision Exp

DeepSeek has introduced V4 Flash Vision Exp, a multimodal addition to its flagship V4 model series. The model is available at launch through the Chinese startup’s paid developer platform, with a free release possible later because DeepSeek has open-sourced many earlier models.

V4 Flash Vision Exp is based on V4 Flash, released in April. In DeepSeek’s tests, it surpassed V4 Flash on text benchmarks except Cybergym, which evaluates software vulnerability discovery, and delivered its largest gains in image analysis. It scored more than 10% higher on two visual tests and outperformed Anthropic PBCC’s Opus 4.8 on ALE and ZeroBench, which measure complex application tasks and difficult image analysis challenges.

DeepSeek has not disclosed the new model’s architecture, but V4 Flash is described as a mixture of experts system with 284 billion parameters and component networks containing 13 billion parameters each. Its KV cache compression techniques, HCA and CSA, are said by DeepMind to reduce computing power for prompts with 1 million tokens by 73%. V4 Flash was trained on 32 trillion tokens using Muon, an algorithm designed to accelerate hidden-layer calibration.