Dev seeks free H200 to scale Luganda chat model Pearl Chat from 8.5M to 500M parameters
weskambale · x · 2026-10-15
Developer weskambale is asking for a spare H200 GPU to train Pearl Chat, a Luganda-language chat model, scaling it from 8.5M to 500M parameters trained purely on Luganda prose and conversations. A screenshot shows Pearl Chat v2 [XL] configured at 508,622,176 total parameters, 100% trainable (CPT/SFT), with 424.7M transformer block params and 83.9M embedding + LM head params — a rare on-the-ground effort at localizing LLMs for an underrepresented language.
More from Fun
- Claude worms across the IoT as civilization's gut microbiome — liminal_bardo · 2026-10-15
- After a 14-hour day with Claude, talking to your family feels like this — rudrank · 2026-10-15
- The real AGI milestone: Windows updates that don't force a restart — chongdashu · 2026-10-15
- Playing video in the terminal: a heavily customized tmux AI-agent desktop goes viral — sull · 2026-10-15
- 3,063 GitHub contributions in a year: dev admits it was just money for AI tokens — KevinNaughtonJr · 2026-10-15
- Dev quip: most AI agents are a far scarier form of Bonzi Buddy — and we're OK with it — JoshuaJBouw · 2026-10-15