How does the Edge0-145B-NVFP4 perform in benchmarks?

#5
by AmixDigital - opened

Hello, I noticed that in your benchmarks, the Edge0-145B-NVFP4 performs very well, but when you say “Edge0-145B,” are you referring to the full-precision version or the NVFP4 version? I was wondering because we don't have access to a version with higher precision.

Edge0 org

“Edge0-145B” refers to the NVFP4 model—it is the exact checkpoint we released. It was trained natively in NVFP4 rather than quantized afterwards, so there is no higher-precision version; the benchmark numbers are from the released weights.

Congratulations on this achievement! It's INCREDIBLE that you've managed to create an NVFP4 model capable of rivaling Qwen 3.8 Flash Next, which offers maximum accuracy.

Actually, only the experts, which are the bulk of the model, are FP4; other parts are BF16 or Q8, just like DeepSeek V4.1 Flash. Experts are insensitive to quantization precision, and DeepSeek just does the same thing, so it's standard practice — nothing incredible about it.

wanglamao changed discussion status to closed

Sign up or log in to comment