AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

DeepSeek launches V4 Flash Vision Exp multimodal model, beating Opus 4.8 on visual tests

DeepSeek has introduced V4 Flash Vision Exp, a multimodal LLM that outperformed Anthropic's Opus 4.8 on two image benchmarks, with a paid platform launch and possible open-source release later.

AI News Feed · 2026-08-22 08:47

V4 Flash Vision Exp is based on V4 Flash, an algorithm DeepSeek released in April. The company said it may later open-source a free version, as it has done with many earlier models. Across seven text-based benchmarks, V4 Flash Vision Exp beat its predecessor on all but one, Cybergym, which tests vulnerability discovery.

The largest gains came in image analysis. DeepSeek measured the model’s visual processing with four benchmarks and said it scored more than 10 percent higher on two tests. On ALE, which contains more than 1,000 multi-step tasks involving interacting with applications, writing code and interpreting media files, and on ZeroBench, a set of 100 deliberately difficult image-analysis tasks, V4 Flash Vision Exp beat Opus 4.8.

DeepSeek has not disclosed V4 Flash Vision Exp’s architecture. Its Hugging Face page, however, describes the underlying V4 Flash model: a mixture-of-experts design with 284 billion total parameters, divided into neural networks of 13 billion parameters each. For a given prompt, the model activates only the most suitable expert network, reducing hardware use compared with running the full model.

To compress the KV cache that stores information for generating answers, V4 Flash uses techniques called HCA and CSA. DeepSeek said these reduce the compute needed to process prompts with 1 million tokens by 73 percent. V4 Flash was trained on 32 trillion tokens using an algorithm called Muon, which shortens the time needed to calibrate the model’s hidden layers.

DeepSeek also released V4 Pro in April, a larger model with more than five times V4 Flash’s parameter count. Since V4 Flash Vision Exp is derived from V4 Flash, the company may eventually build specialized models on V4 Pro for tasks such as image analysis.