Apple's STARFlow2 description puts understanding, reasoning and interleaved text-image generation in one generation design. The architecture uses flows that share a language model's left-to-right order, causal mask and KV cache. Apple reports strong results for image generation and multimodal understanding but gives no scores, code release or weights-release plan.
Apple's machine learning group introduced STARFlow2, which handles understanding, reasoning and the generation of interleaved text-image sequences within one causal design. The description places those functions in a shared generation stream rather than as disconnected tools.[1]
The model uses autoregressive normalising flows that share a language model's left-to-right order, causal mask and KV cache. The technical description joins a frozen vision-language-model stream with a TARFlow stream on the Pretzel architecture.[1]
Apple reports strong results on image-generation and multimodal-understanding benchmarks but provides no scores. The description also does not say that code or weights will be released. It supplies an architectural account, not a package for externally reproducible comparison. The published material explains how the shared causal stream is built but does not give scores that measure the size of its performance claims. With no stated code or weights release, it also leaves no specified route for outsiders to run the same design. Because the single-stream design uses the same left-to-right order, causal mask and KV cache for text and image sequences, it may contribute to handling those functions together. But because Apple provides no scores, it cannot be determined whether the reported strong results come from that design or from undisclosed evaluation conditions.[1]
