Downloads · 30 days
2
2% of all-time downloads
Qwen3-coder-480B-fused-bf16
anirudmohan26/Qwen3-coder-480B-fused-bf16
Qwen3-coder-480B-fused-bf16 is a machine learning model from anirudmohan26. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Offline RMSNorm fusion: inputlayernorm gamma folded into q/k/vproj weights, gamma reset to 1. Mathematically equivalent to the base model (validated on Qwen3-0.6B: fp32 maxdiff ≤ 7.5e-6; bf16 logits within noise).
All-time downloads
97
Public
Parameters
480B
960 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors960 GB · 100%
At a glance
- Model type
- qwen3_moe
- Access
- Public
- Created
- May 29, 2026
- Updated
- May 30, 2026
- SHA
- d6fb71df
Base models
Model card
From the Hugging Face model README
Qwen3-Coder-480B-A35B — RMSNorm-fused (bf16)
Offline RMSNorm fusion: input_layernorm gamma folded into q/k/v_proj weights,
gamma reset to 1. Mathematically equivalent to the base model (validated on
Qwen3-0.6B: fp32 max_diff ≤ 7.5e-6; bf16 logits within noise).
⚠️ This checkpoint alone is NOT faster. It is the input to the runtime fused
CUDA kernel. To get the speedup, serve with patch_qwen3_model(variant="V1")
from the qwen3-rmsnorm-fusion repo. Primary purpose: artifact for downstream
quantization.
⚠️ Quantization + kernel integration is unresolved: the V1 kernel does a plain matmul on bf16 combined-QKV weights and does not yet consume quantized (int4/int8/fp8) weights — decide how the fused QKV is handled when picking a quant method.
- Type
- qwen3_moe
- Created
- May 29, 2026
- Updated
- May 30, 2026