Aurora2-400M-A20M-QAT: a 0.5-bit MoE checkpoint shrunk from 799 MB to 27.5 MB
Tall_Abrocoma_3533 · reddit · 2026-11-23
The author releases a quantization-aware-trained version of Aurora2-400M-A20M using LittleBit-2 binary low-rank quantization with distillation from the original model. The routed experts (98% of parameters) are stored at roughly 0.5 bits per weight, shrinking the checkpoint from 799 MB in FP16 to 27.5 MB.
The quality cost is large: the Intelligence Index drops from 21.38 to 10.83. Arithmetic holds, PIQA stays above chance, while HellaSwag and ARC-Challenge fall near chance.
The compression is storage-only: without dedicated binary kernels, the loader expands weights back to FP32, so memory and speed are no better than the base model. It targets research on extreme low-bit MoE quantization, not a drop-in replacement.
More from Models
- Kimi K3.1 Pro on Cerebras hits 1375 TPS after 594 agent-run experiments — hyperparticle · 2026-11-23
- Perplexity is a floor metric, not a magic bullet, argues researcher — eliebakouch · 2026-11-23
- Leaked projection: OpenAI internal model at 0.732 perplexity, automated AI research by May 2027 — rickasaurus · 2026-11-23
- Aurora2-400M-A20M-SFT: chat fine-tune of a sparse MoE with only 19M active parameters — Tall_Abrocoma_3533 · 2026-11-23
- Labs sharing base-model perplexity metrics could mirror post-training evals — eliebakouch · 2026-11-23
- ChatGPT keeps promising not to violate your brand guidelines, and users are annoyed — chickefitz · 2026-11-23