The new Apache 2.0 model enables developers to convert LLM outputs into interactive visual components without relying on proprietary frontier APIs.
WARSAW, Poland — 02.07.2026 — Mobile Reality (custom software development from Poland) has released mdma-gemma4-26b-dsl-unsloth-v1 on Hugging Face, an Apache 2.0 language model that powers the company's MDMA framework for transforming plain LLM text into interactive user interface components. The release allows organizations to self-host generative UI capabilities using open-source infrastructure rather than relying on expensive proprietary frontier APIs. By fine-tuning a mid-size mixture-of-experts model for domain-specific language tasks, Mobile Reality aims to make interactive AI interfaces practical for enterprises with strict data sovereignty requirements.
The model is built on Google's Gemma 4 26B-A4B mixture-of-experts architecture and was fine-tuned using Unsloth's 16-bit LoRA on a single NVIDIA A100 80GB GPU. Mobile Reality merged the adapter weights into the base model and serves the resulting checkpoint locally with vLLM int8 quantization. The final model offers a 128,000-token context window, native tool-calling support, and agentic behaviors optimized for complex workflows. Specializing in MDMA-IL, the model translates high-level interface instructions into valid MDMA YAML components nested inside fenced code blocks within standard Markdown documents. This lightweight training pipeline demonstrates that customized generative UI models can be produced without access to large compute clusters.
MDMA, short for Markdown Document with Mounted Applications, is the open-source generative UI engine that interprets the model's structured outputs. Traditional generative UI approaches typically require models to emit raw JSON or HTML, which can corrupt the entire interface if the output drifts or truncates. MDMA replaces this fragility with markdown-native fenced blocks that allow any renderer to fall back to readable text when components are incomplete. A deterministic validator and automated fixer enforce structural correctness, enabling the system to achieve a 98% reliability rate in production. The project's modular ecosystem includes the @mobile-reality/mdma-spec definition, the mdma-parser and mdma-runtime engines, a mdma-renderer-react package for front-end integration, and a dedicated mdma-mcp server that exposes the engine to Claude Desktop, Cursor, and VS Code through the Model Context Protocol. Developers can evaluate changes using mdma-evals, a suite that covers more than 70 distinct test cases. All MDMA libraries are published under the MIT license and maintained in the Mobile Reality GitHub repository.
Commercially, Mobile Reality applies the technology through its AI Headless Forms product, which converts rigid multi-step enterprise forms into dynamic conversational experiences. Organizations in regulated sectors such as FinTech and PropTech frequently prohibit the transmission of customer personally identifiable information to external API providers. Self-hosting the 26B parameter model behind a private endpoint eliminates that dependency while preserving the interactive features normally associated with frontier models. The system renders forms, tables, approval gates, charts, callouts, webhooks, and reasoning blocks in real time without exposing sensitive data to third parties. Because the underlying inference runs on commodity hardware using quantized weights, operational costs drop substantially compared to per-token pricing from closed-platform providers.
"Training a 26-billion-parameter mixture-of-experts model on a single A100 80GB GPU and serving it locally with vLLM int8 demonstrates that highly capable generative UI inference is accessible to most engineering teams" said Marcin Sadowski, Chief Technology Officer at Mobile Reality. "This release gives developers a fully open, self-hostable path to agentic interfaces with 128K context and native tool-calling, and it validates that the first production version of our specialized model stack is ready for broad use."
"Enterprises in regulated sectors have been forced to choose between expensive proprietary APIs and sacrificing data sovereignty," said Matt Sadowski, Chief Executive Officer at Mobile Reality. "By open-sourcing this model under Apache 2.0, we eliminate that trade-off and make dynamic, conversational interfaces—like our AI Headless Forms—a practical reality for FinTech and PropTech organizations that require absolute control over customer data while still delivering modern user experiences."
About Mobile Reality
Mobile Reality is an AI automation agency and end-to-end product software development company headquartered at Grzybowska 62 in Warsaw, Poland. Founded in 2016 by Matt Sadowski and Marcin Sadowski, the organization combines deep expertise in full-stack JavaScript, React, Node.js, and AWS with a highly AI-driven approach to building secure, scalable digital products. The company holds a 5.0 out of 5 rating on Clutch across 17 client reviews and specializes in ROI-driven AI implementations, on-premises LLM deployments, and high-efficiency software architectures for regulated industries.
