Open-weight model with 20 billion parameters competes with closed systems in corporate slide design

The presentations produced by leading models may appear polished at first glance, but repeated use quickly reveals a uniform visual template that is hard to break away from, as those models always revert to their default style regardless of how precise the textual prompts are. The issue becomes more complex in business environments that require strict adherence to a specific visual identity or that prohibit sensitive data and board-meeting minutes from being sent to external cloud servers, making reliance on closed commercial systems a risky corporate choice with accumulating costs.

The technical bottleneck is that small open-weight models, the natural candidate for on-premise hosting within companies, traditionally fail to generate visually sound slides. Building a slide forces the model to handle two contradictory levels of decision-making at once:editorial estimationto decide what to write and what to delete, andgeometric calculationfor text positions, font sizes and element coordinates without a prior visual preview of the output, which in small models leads to overlapping boxes and words spilling outside slide boundaries.

IBM researchers addressed this dilemma in the "Palette" project by simulating human work partitioning, training an open model weighing twenty billion parameters with the LoRA technique to share two separate tasks. In the "designer" role, the model reads the draft and produces a structural brief in data format that specifies colour schemes, fonts and layout options; in the "programmer" role, the same model writes JavaScript code through a specialised library that draws each slide individually. This decomposition allowed code-generation processes to run in parallel for each slide, increasing the programming stage speed by roughly six to seven times compared with sequential generation.

The team generated synthetic data comprising about a thousand presentations with roughly ten thousand slides, embedding an alternative logical-reasoning path that explains why certain design choices were rejected and others preferred. The experiment highlighted a pivotal lesson in model engineering: direct textual instructions failed to prevent hallucinations when inputs were sparse, whereas providing complete reference presentations within the context calibrated the model’s accuracy, as the system replicated visual references with a strictness that surpassed its response to abstract commands.

To ensure the safety of corporate output, the researchers did not leave the results to chance; each slide passes through an automated inspection environment that measures dimensional geometry and detects unintended overlap or cropping before being handed to an editor model that rewrites the code.Programmatic edits are only accepted if the automated check demonstrates an improvement in slide quality, a mechanism that alone added three and a half points on average to the system’s rating. The system also supported direct textual instructions in double quotes that allow the user to convert tables into timelines or impose a direct, concise tone without resetting the model.

Test results on the specialised "Present Bench" metric, which evaluates five criteria, design, completeness, accuracy, source conformity and others, showed Palette ranking first in the economics and education categories and second overall, trailing the "Notebook LM" system by only a tenth of a point (59.7 versus 60.8), outperforming composite commercial systems such as "Manus 1.6" and five other competing solutions despite using a limited-size open model.

This shift outlines clear implications for tech teams and data departments across the Gulf, Egypt and the regional ecosystem; banking sectors and government bodies that manage sovereign assets face strict governance constraints that forbid sending financial performance reports and investment strategies through cross-border cloud APIs. The approach gives regional institutions the ability to host medium-weight twenty-billion-parameter models on modest internal servers and adapt the training adapter to enforce corporate visual identity and approved workflow templates, instead of draining budgets on cloud-model subscriptions and token consumption without a genuine guarantee of output consistency.

The significance of this development goes beyond office-presentation production to confirm the viability of guided systems that embed specialised models within programmatic verification frameworks. With developers moving toward using geometric inspection metrics as rewards in reinforcement-learning model training, the operational gap for companies seeking to own their intelligent software narrows;the real bet in corporate production no longer rests on model size alone, but on the rigor of the surrounding inspection environmentthat ensures result safety and reliability before they reach the decision-making table.