Apple introduces MANZANO — a unified multimodal model that balances image understanding and generation

The trade-off problem in multimodal models

Multimodal models that simultaneously understand and generate images have suffered from a fundamental trade-off for years. Systems optimized for image understanding — typically relying on continuous embeddings — excel at describing content but struggle to generate new images. Conversely, models that generate images well usually use discrete tokens and an autoregressive architecture that struggles with detailed descriptions. Merging both worlds into a single model has until now meant sacrificing quality on at least one side.

At ICLR 2026, whose program Apple Machine Learning Research published, Apple’s team introduced MANZANO — a framework that attempts to close that gap. According to the announcement, MANZANO offers a unified architecture that simultaneously balances image understanding and generation within a single model, without the need for separate systems for each task.

Hybrid vision tokenizer and dual adapters

The key technical innovation in MANZANO is its hybrid vision tokenizer. Rather than offering exclusively continuous embeddings — preferred by understanding models — or exclusively discrete tokens — preferred by generative models — the tokenizer produces both representations from the same input signal. Continuous embeddings serve as rich semantic input for image understanding, while discrete tokens are used in autoregressive decoding during generation.

Above that shared encoder, MANZANO uses two specialized adapters — one for each type of task. This approach, which Apple describes in its announcement as “shared encoder, dual adapters,” means the model shares the majority of its parameters and representations but has specialized heads at the top trained for different objectives. The result, according to the authors, is a reduction in the trade-off between the two tasks compared to existing unified approaches.

Why it matters

If the results hold up in broader practice and on independent benchmarks, MANZANO has the potential to change how multimodal applications are built. Developers today often combine two separate models — for example Claude or GPT-4V for understanding and Stable Diffusion or Flux for generation — which means double serving costs, a more complex pipeline, and harder maintenance. A unified model like MANZANO allows the same system to follow a conversation, understand an attached image, and generate a new one, without switching context between models.

Such an architecture particularly opens the door to interactive scenarios like image editing through conversation, where the user describes desired changes in natural language and the model understands both the image and the instruction and generates a new version. Apple has not released MANZANO weights nor announced when the feature might appear in products, but the ICLR publication signals the direction of Apple’s research and the potential for integration into future versions of Siri, Final Cut Pro, or generative tools in iOS. For the broader community, MANZANO is a valuable reference point showing that unifying understanding and generation does not necessarily mean a loss of quality.