DeepSeek launches multimodal model matching Anthropic's Opus 4.8

What happened

DeepSeek unveiled V4 Flash Vision Exp, a multimodal model derived from its V4 Flash algorithm released in April. The new model beat Anthropic's Opus 4.8 on two visual benchmarks—ALE, which contains more than 1,000 multi-step tasks requiring code writing and media interpretation, and ZeroBench, which has 100 image analysis tasks designed to challenge frontier LLMs. V4 Flash Vision Exp outperformed V4 Flash across all seven text-based benchmarks except Cybergym, which evaluates software vulnerability detection.

Why it matters

The debut signals DeepSeek's competitive positioning in multimodal AI at a time when enterprises increasingly require models that handle both text and images. V4 Flash Vision Exp's stronger performance on image analysis—scoring more than 10% higher on two visual benchmarks—shows the company is closing gaps with established competitors. The model's foundation, V4 Flash, uses a mixture-of-experts architecture with 284 billion parameters that activates only the neural network best suited to each prompt, reducing hardware demands compared to activating the entire model.

What to watch

V4 Flash Vision Exp is currently available only via DeepSeek's paid developer platform, though the company may release a free version later, consistent with its practice of open-sourcing earlier models. DeepSeek has not disclosed V4 Flash Vision Exp's architecture details. V4 Flash was trained on 32 trillion tokens of data and uses compression techniques (HCA and CSA) that reduce computing power needed to process 1 million-token prompts by 73%.