Uganda team ships 350M-param Luganda chat model, but compute is exhausted
weskambale · x · 2026-11-14
The AI for Uganda team released pearl-chat-350m-it, a strictly monolingual 350M-parameter Luganda conversational model, a big step up from their earlier 46M and 85M checkpoints.
- Multi-stage training: broad pretraining, WSD Uganda knowledge annealing, then length-bucketed SFT on 14,809 single- and multi-turn conversations
- Built with JAX + Flax NNX, using grouped-query and sliding-window attention; Apache-2.0, gated weights on Hugging Face
- The team admits some SFT overfitting near the end and says they've burned through all their compute
More from Models
- LLMs playing AoE2 keep mistaking gold piles for berries, showing vision is unsolved — max_paperclips · 2026-11-14
- Claude MCP goes offline in rare outage, as blogger flags too many failures to be coincidence — AiBreakfast · 2026-11-14
- OpenAI $200 plan drops to $100 on Nov 20; user races to burn 4 banked Codex resets — koltregaskes · 2026-11-14
- SenseNova 6.8 Flash ships long-term memory, multi-step workflows and editable 3D output — FellMentKE · 2026-11-14
- Krishnan: With infinite compute, you could just let AI hallucinate freely — krishnan · 2026-11-14
- GPU forensics suggest DeepSeek's real compute far exceeds disclosed figures — teortaxesTex · 2026-11-14