Full AI Chat and Image Generation on a 1990 Tandy 286: 9s Draw Times, Open Source
jacobpederson · reddit · 2026-10-01
A developer got full AI chat and image generation running on a 1990 Tandy 1000 TL/3 (286) via the open-source DeskMind (GPLv3). The DOS program talks WiFi (PicoMEM 2 + mTCP) to a Python server driving Qwen3.8-27B on a 5090 and Krea 2 on a 4090. The 286 only sees plain text and raw video-memory-ready pixels — no JSON, base64 or PNG. Qwen wraps draw requests in <draw> tags caught mid-stream; Krea 2 renders and dithers in 9s. Streaming cleanup strips reasoning and Markdown, converts to CP437, and merges tokens into 48-char lines. Includes a Dither Lab GUI and per-task reasoning budgets.
Related event: Developer Runs AI Chat and Image Generation on a 40-Year-Old 286 PC(2 posts)→
More from Fun
- Believing AI is conscious is a potent cognitive security vulnerability, says Yacine — yacineMTB · 2026-10-01
- Using AI Classifiers to DJ: Scoring Which Track Mixes Into the Next — Aizkmusic · 2026-10-01
- Mocking Sorkin's OpenAI movie: "What if we dubbed it into French?" — dbasch · 2026-10-01
- While OpenAI and Meta race ahead, the new Siri takes 30s then says "I didn't quite catch that" — firstadopter · 2026-10-01
- Cached or calculated? The multiplication flashcards debate mirrors the LLM memory question — dbasch · 2026-10-01
- New Bay Area status symbol: being visibly stressed about lack of compute — matt_slotnick · 2026-10-01