Presenting our ICLR Spotlight at Singapore Expo! Hall 3, 3 PM.
Pretrained Ternary LLMs (TriLMs) outperform or match Float/Quant models per bit at >1B scale. Our 3.9B TriLM matches 3.9B FloatLM but is tiny (fewer bits than 830M FloatLM)!
Link: iclr.cc/virtual/2025...
#ICLR #AI #LLM
Pretrained Ternary LLMs (TriLMs) outperform or match Float/Quant models per bit at >1B scale. Our 3.9B TriLM matches 3.9B FloatLM but is tiny (fewer bits than 830M FloatLM)!
Link: iclr.cc/virtual/2025...
#ICLR #AI #LLM
ICLR Poster Surprising Effectiveness of pretraining Ternary Language Model at ScaleICLR 2025
iclr.cc
April 26, 2025 at 7:01 AM
Presenting our ICLR Spotlight at Singapore Expo! Hall 3, 3 PM.
Pretrained Ternary LLMs (TriLMs) outperform or match Float/Quant models per bit at >1B scale. Our 3.9B TriLM matches 3.9B FloatLM but is tiny (fewer bits than 830M FloatLM)!
Link: iclr.cc/virtual/2025...
#ICLR #AI #LLM
Pretrained Ternary LLMs (TriLMs) outperform or match Float/Quant models per bit at >1B scale. Our 3.9B TriLM matches 3.9B FloatLM but is tiny (fewer bits than 830M FloatLM)!
Link: iclr.cc/virtual/2025...
#ICLR #AI #LLM