Research
Quantization
Smaller models that preserve the quality of their answers.
BBQ-FP4 is our own 4-bit quantization. It cuts model-weight memory by 3.2x to 3.8x while matching or exceeding full precision on average benchmark recovery.
3.2–3.8x
smaller model-weight footprint than bf16
101.4%
average recovery vs bf16 on Qwen3-30B-A3B (101.3% on Qwen3-Next-80B-A3B)
2.67x
decode throughput at best: 6,151 tok/s vs 2,304 in bf16 on Qwen3-30B-A3B
The weights, to scale.
Model-weight memory in bf16 and in BBQ-FP4. One block is 8 GB.
Qwen3-30B-A3B
Qwen3-Next-80B-A3B
Against another 4-bit build of the same size.
Both 4-bit builds use about the same memory. Higher recovery and lower KLD mean a closer match to bf16.
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 is for BBQ-FP4-S; BBQ-FP4-L reached 100.7%.
The code is not public yet.