GLiNER
Distilling a 287M model from Claude worked — except for that last 5% of accuracy
A 287M model distilled from Claude handled 5M court rulings — but missed the last 5% of accuracy. The real question: can AI be made affordable?
Back to home
model distillation
2 articles tagged with this topic
Xyntetik-Kvist-14B
30B Model Halved to 14B, 95% Agent Capability Kept — Local on One Consumer GPU
A solo developer cut a 30B Muse-Glimmer model in half and distilled it to a 14B Xyntetik-Kvist-14B. 57 of 60 tool-call tasks pass; quantized, it fits