More Qwen MoE capacity per GPU, without giving up model quality.

BBQ-FP4 reduces model-weight memory by 3.2x to 3.8x while matching or exceeding full-precision average benchmark recovery.

3.2x – 3.8x

smaller model-weight footprint

16 GB vs 57 GB on Qwen3-30B-A3B

40 GB vs 152 GB on Qwen3-Next-80B-A3B

101.4% / 101.3%

average recovery

Across the two Qwen MoE models

versus bf16 reference

up to 2.67x

decode throughput

6,151 tok/s on Qwen3-30B-A3B

vs 2,304 tok/s in bf16

Benchmark evidence: BBQ vs. external 4-bit baseline

Both comparison models use approximately the same 4-bit footprint. Higher recovery and lower KLD indicate a closer match to bf16 behavior.

Metric
bf16
BBQ-FP4
External w4a16
BBQ edge
Qwen3-30B-A3B
Avg. recovery
100.0%
101.4%
99.7%
+1.7 pts
Worst-case recovery
100.0%
98.5%*
98.1%
+0.4 pts
WikiText KLD
0
0.0345
0.0733
2.1x lower
Weight memory
57 GB
16 GB
~16 GB
same footprint
Qwen3-Next-80B-A3B
Avg. recovery
100.0%
101.3%
97.8%
+3.5 pts
Worst-case recovery
100.0%
94.8%
93.1%
+1.7 pts
WikiText KLD
0
0.0285
0.0329
13% lower
Weight memory
152 GB
40 GB
~41 GB
~same footprint
What this means for cloud providers

Increase deployable model capacity within existing GPU memory envelopes. The same compression class becomes a quality differentiator, not just a cost lever.

Where BBQ is strongest

Quality-sensitive MoE serving where benchmark regressions, distribution drift, and GPU memory ceilings matter as much as raw model compression.

Methodology: Post-training FP4 quantization of Qwen3 MoE models. Recovery and memory savings are relative to bf16. External comparison values marked ~ are extrapolated, not directly benchmarked. Standard benchmarks are proxies and may not represent all production workloads.

* Qwen3-30B-A3B worst-case recovery: BBQ-FP4-S; BBQ-FP4-L achieved 100.7%.