The authors formalize MoE weight precision as two rate functionals: a resident rate (what is stored) and a decode rate (what is read per step).
Two-rate precision allocation cuts MoE memory footprint and decode traffic
DuoRate stores weights once and adapts read precision per workload, matching 4-bit accuracy at 2.9 bpw
Chinese Tech
Mengqi Guo · Zhe Wei · Yuan Yuan · Jiunn Bin Lim · Boyi Pan · Michael Bi Mi
Huawei Technologies Ltd.
Research Digest··3 min read
95× for BitsMoE).
Why this paper
From Huawei Technologies Ltd.
In one line
DuoRate stores MoE weights once and allocates per-expert read precision per workload, matching 4-bit formats at 2.9 bpw on Qwen3-30B-A3B.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§