DeskMind
A DOS-native program created by developer Rowan Underwood that bridges modern large-model inference capabilities to 1990s 286-architecture PCs through protocol down-conversion, enabling AI chat and image generation on legacy hardware.
Timeline (last 90 days)
Developer Rowan Underwood used a system called DeskMind to run AI chat and image generation on a 1990s 286 Tandy 1000 TL/3 computer
DeskMind is a native program running under DOS; the 286 connects to WiFi via a PicoMEM 2 network card with the mTCP protocol stack and communicates with a lightweight Python server on a PC
The server pre-processes all data into plain text lines and image data that can be copied directly into video memory, so the 286 never needs to handle JSON, base64, or PNG
The project uses a system prompt to instruct Qwen to wrap drawing requests in <draw>...</draw> tags; the server captures these tags in the output stream to trigger image generation, rather than relying on function calling
In the author's tests, from pressing Enter on a drawing request to a thumbnail appearing in the chat window takes approximately only 9 seconds
The system prompt informs the model that it is running on an 80-column, 16-color 286, causing Qwen to proactively keep responses brief and use plain ASCII characters
On the server side, Unicode is converted to the CP437 code page, and scattered tokens are merged into lines of approximately 48 characters to prevent the 286 from re-rendering the screen token by token
The server GUI has a built-in Dither Lab that supports multiple dithering algorithms including Floyd-Steinberg, Atkinson, Bayer, and Yliluoma
The DeskMind project code has been open-sourced under GPLv3 and is hosted on GitHub
All Facts (9)
Developer Rowan Underwood used a system called DeskMind to run AI chat and image generation on a 1990s 286 Tandy 1000 TL/3 computer
50%UnverifiedDeskMind is a native program running under DOS; the 286 connects to WiFi via a PicoMEM 2 network card with the mTCP protocol stack and communicates with a lightweight Python server on a PC
50%UnverifiedThe server pre-processes all data into plain text lines and image data that can be copied directly into video memory, so the 286 never needs to handle JSON, base64, or PNG
50%UnverifiedThe project uses a system prompt to instruct Qwen to wrap drawing requests in <draw>...</draw> tags; the server captures these tags in the output stream to trigger image generation, rather than relying on function calling
50%UnverifiedIn the author's tests, from pressing Enter on a drawing request to a thumbnail appearing in the chat window takes approximately only 9 seconds
50%UnverifiedThe system prompt informs the model that it is running on an 80-column, 16-color 286, causing Qwen to proactively keep responses brief and use plain ASCII characters
50%UnverifiedOn the server side, Unicode is converted to the CP437 code page, and scattered tokens are merged into lines of approximately 48 characters to prevent the 286 from re-rendering the screen token by token
50%UnverifiedThe server GUI has a built-in Dither Lab that supports multiple dithering algorithms including Floyd-Steinberg, Atkinson, Bayer, and Yliluoma
50%UnverifiedThe DeskMind project code has been open-sourced under GPLv3 and is hosted on GitHub
50%