What I just uploaded
I just put Qwen3-Coder-30B-A3B-Instruct-NVFP4 on Hugging Face. It is an NVFP4 quantization of Qwen/Qwen3-Coder-30B-A3B-Instruct. The base is a mixture of experts model with 30B parameters and 3B active per token. The model is at https://huggingface.co/kleinpanic93/Qwen3-Coder-30B-A3B-Instruct-NVFP4. The license is Apache 2.0, the same as the base.
How I made it
I used post-training quantization with NVIDIA Model Optimizer 0.41.0 and NVFP4_DEFAULT_CFG. It ran on the GB10 in my DGX Spark with a script called spark-maker v3. The config sets weights and activations to 4-bit NV floating point, group size 16. The KV cache is FP8. The lm_head and all the MoE router and gate layers are left out, 48 layers in total.
ModelOpt 0.41.0 cannot export Qwen3MoeExperts natively. So I saved with save_pretrained and injected the quantization_config by hand.
Two uploads
The first upload used synthetic random calibration data. About an hour and a half later I replaced the weights with a requantization calibrated on real allenai/c4 text. The second run took 267 seconds.
The model card was not updated. It still says synthetic calibration. The provenance JSON in the repo and the last commit message have the right story: C4.
What I am not claiming
I make no claim about size. I also make no claim about how a loader treats the checkpoint.
There are no eval numbers for this model. The card only reports how long the quantization took.
What is next
If you load it, tell me what you see. I want to know.
P.S. Thanks for reading. Consider subbing to the RSS feed.
Klein Panic