IT之家 · 智能时代T2·78 pts
- Shared loop module reuses layers 4–27 twice, yielding 56 decoder layer computations per forward pass
- 4-bit weights are about 4.9 GB; 256K-context inference needs roughly 15 GiB of memory
- Scores 41.36 on NL2Repo and 65.31 on AIDABench, competitive with Qwen3.8-27B
- Official AWQ 4-bit quantization ships day one alongside open weights and code