A Transformer with around 30 million parameters, built around a shared vocabulary of 50,520 tokens and capable of generating text, reconstructing images, or producing rough visual sketches from textual descriptions. It’s not a finished product, but with Supra-A2A-Nano-Exp, SupraLabs signs a conceptual experiment that fascinates: every modality – text, image, video – becomes a sequence of tokens, and everything else is simple next-token prediction.

The key lies in the architecture. There is no separate vision encoder, no diffusion model, no cross-attention modules between heterogeneous streams. Images are broken into patches by a VQ-VAE with 256 codes, producing 8×8 grids of tokens for a 64×64 pixel input. Videos are sequences of frames treated the same way. The text side uses a GPT-2 style BPE tokenizer with 50,264 tokens, joined by the 256 visual codes into a unified vocabulary. Special markers (<TEXT>, <IMAGE>, <VIDEO>, <FRAME>) bound modalities, but for the model it’s all the same language.

The backbone is a GPT-like Transformer with 4 layers, embedding size 256, maximum context of 384 tokens, likely 4 attention heads and an MLP with 4× expansion. Weights are distributed in FP32 safetensors format. The entire pipeline runs with a few lines of Python: import torch, transformers, safetensors, Pillow, numpy. The “text2image” mode produces an image from a textual prompt wrapped in the stream, for example “<TEXT>a red square</TEXT><IMAGE>”.

Beyond the technical curiosity, this project shines a light on trade-offs that anyone evaluating on-premise deployment should keep in mind. The idea of replacing complex multimodal pipelines – encoders, decoders, diffusion modules, orchestration between components – with a single Transformer trained to predict the next token over an enlarged vocabulary slashes architectural complexity. For local workloads, that means fewer software dependencies, lower memory demands from keeping separate models, and the ability to run inference on modest hardware. The ~30M parameter FP32 network can run on CPU or low-VRAM GPUs, making fully self-hosted processing feasible even without high-end accelerators.

Of course, clear limitations exist. The model is small, visual resolution is low and abstract, it lacks any RLHF or instruction tuning, and the 384-token context is tight. It’s a research prototype, not a production solution. Yet it represents a direction that on-premise stack designers should observe: multimodality doesn’t have to be an assembly of specialized bricks. The “everything is tokens” approach radically simplifies inference, reduces integration points, and can contain Total Cost of Ownership (TCO) over the long term.

More broadly, the model reminds us that the hypertrophy of current architectures isn’t the only path. While large labs push systems with hundreds of billions of parameters and complex orchestrators, Supra-A2A-Nano-Exp acts as a prod: it draws attention back to conceptual simplicity and deployment practicality. For an enterprise that must run local models shielded from cloud and external APIs, the question isn’t just “which model is more accurate?” but also “how much infrastructure is really needed?”. The SupraLabs experiment doesn’t provide ready-made answers, but it poses the question in the cleanest possible way.

The Any2Any family, just inaugurated, is an open workshop. The repository offers a nano model and tools to run text generation, chat, image reconstruction, and text-to-image. The community is invited to experiment, but the most immediate value may lie in triggering a reflection: if modalities are truly reducible to tokens, then on-premise deployment can become much leaner. And the adoption of Large Language Models in contexts with data sovereignty and operational control constraints could accelerate precisely where simplicity pays.