A Tandy 1000 TL/3 with a 286 processor can chat and generate images without ever processing a single token. The credit goes not to the 1990 machine but to an architecture that shifts the workload to a local server. DeskMind, a native DOS program, talks over WiFi through a PicoMEM 2 card using mTCP and connects to a small Python server on a PC. That server runs Qwen3.8-27B with NInfer on an RTX 5090 and Krea 2 with ComfyUI on an RTX 4090. The 286 never sees JSON, base64, or PNG: it receives plain text lines and images ready to be copied into video memory.

Image generation happens without tool calling. The system prompt tells Qwen to wrap a picture request in <draw>...</draw>. The server catches the tag mid-stream, runs Krea 2, applies dithering, and returns a picture ready line. The request 'Draw me a 286 AI logo' takes about 9 seconds from Enter to a thumbnail in the chat. Qwen Vision receives both the original and the 16-colour dithered version, so when you ask why the sky looks striped, it has real context. The latest picture stays in context for follow-up questions.

The adaptation layer is the most interesting part. The server strips reasoning, removes Markdown on the fly, converts Unicode to code page 437, and merges smaller tokens into lines of about 48 characters, keeping the 286 from redrawing the screen for every token. The system knows it lives in an 80-column, 16-colour environment: it keeps answers short and plain ASCII, and when asked about games it suggests Wolfenstein 3D or Commander Keen. Per-request reasoning effort is low for chat, with replies starting in about 2 seconds, and medium for rewriting image prompts. The rewrite appears in an edit box on the Tandy before drawing, and the rules themselves can be edited from the Tandy.

There is also a Dither Lab in the server GUI: Floyd-Steinberg, Atkinson, Bayer, Yliluoma, and others, previewed at the Tandy's real aspect ratio. Krea 2 generates at 1024x768 in about 10 seconds with 8 steps, but the target is 640x200. Sending a 64,000-byte image over the PicoMEM WiFi card takes about 1 second, with observed throughput between 56 and 79 KB/s. In other words, in the image generation path, the network transfer weighs more than local compute, despite the modern GPU.

This project is not a simple tribute to retrocomputing. It is a practical demonstration of how a terminal-server architecture can separate inference from presentation. The Tandy acts as a thin client: it does not need to run the model, decode JSON or PNG, or manage high-resolution rendering. All computational load stays on the local server. For anyone evaluating on-premise deployment, this is a structural signal: hardware cost shifts from the client fleet to the central node, which must host current GPUs and manage power and cooling. Clients can remain obsolete as long as they can receive text and images in a compatible format. The local nature of the server also means data does not leave the home or corporate network, but it requires expertise to maintain the stack. For those weighing on-premise trade-offs, AI-RADAR offers analytical frameworks at /llm-onpremise without prescribing a single path.

Dithering, often dismissed as a cosmetic detail, becomes a real compatibility layer here: it lets a 16-colour display show an image generated at 1024x768 without transmitting or processing it locally. The same applies to stream cleanup: turning LLM output into code page 437-friendly lines is not a gimmick, it is a presentation protocol. The project is distributed under GPLv3 on GitHub, which makes the whole experiment reproducible and inspectable.