The number that matters is 284 billion: the parameters inside DeepSeek's V4 Flash mixture-of-experts backbone, of which only about 13 billion activate for any single prompt. That backbone is the foundation of V4 Flash Vision Exp, the multimodal agent model DeepSeek debuted Friday on its paid developer platform. The "Exp" tag is doing honest work: the company has published no architecture notes for the new model, and its usual open-weights release is unconfirmed, a notable reversal for a lab whose drops typically lead with the checkpoint.
The vendor's benchmark sheet is specific. Against its own predecessor, Vision Exp wins six of seven text benchmarks, the sole loss coming on Cybergym, which measures software-vulnerability discovery. The gains concentrate where the name says they should: across four visual benchmarks the model improves on V4 Flash by more than 10 percent on two, and SiliconANGLE reports it also beats Anthropic's Opus 4.8 on ALE, a suite of more than 1,000 multi-step tasks that require interacting with applications, writing code, and interpreting media, and on ZeroBench, 100 image-analysis tasks designed to be hard for frontier models. None of this has been independently evaluated yet; treat the charts as claims until third-party runs land.
Underneath sits the engineering the V4 Flash model card documented in April: HCA and CSA, two KV-cache compression techniques that cut the compute needed for million-token prompts by roughly three quarters, training over 32 trillion tokens, and the Muon optimizer to speed hidden-layer calibration. DeepSeek's larger V4 Pro, with more than five times the parameters, is the obvious candidate for the same vision treatment if Vision Exp's pattern holds.
For builders, the story is the price-performance direction of computer-use agents. A screen-reading agent model from a lab with DeepSeek's cost structure pressures the assumption that serious visual agent work requires a frontier-flagship budget; Canaltech's coverage describes exactly that use case, tasks solved by looking at screens. Watch two things: whether independent benchmarks confirm the Opus 4.8 comparisons, and whether the weights follow, because a downloadable screen-reader at 13 billion active parameters would move the local-agent tier again, two weeks after Muse Glimmer and Qwen 3.8 did.