Qwen3.6-35B-A3B-sft-15020
Qwen3.6-35B-A3B-sft-15020 is a code generation model obtained by supervised fine-tuning (SFT) of Qwen/Qwen3.6-35B-A3B, specialized for NPU (Ascend) operator kernel generation. Given an operator's functional description and interface definition, the model generates kernel implementations (e.g., AscendC) that run on Ascend NPUs.
Model Summary
| Item | Description |
|---|---|
| Base model | Qwen/Qwen3.6-35B-A3B (MoE model with 35B total / 3B activated parameters) |
| Training method | Supervised fine-tuning (SFT) |
| Training framework | MindSpeed-MM |
| Task | Code generation / NPU kernel generation |
| Evaluation benchmark | NPUKernelBench |
| License | Apache-2.0 |
Quick Start
The model shares the same architecture as Qwen3.6-35B-A3B and can be loaded directly with transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen3.6-35B-A3B-sft-15020"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto",
)
prompt = "Implement an AscendC operator that performs element-wise addition of two input tensors."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
The model can also be served with inference frameworks such as vLLM or SGLang, following the same usage as Qwen3.6-35B-A3B.
Evaluation
The model is evaluated on NPUKernelBench, a benchmark for NPU kernel code generation that measures the correctness (compilation success and numerical accuracy) and performance of generated operator kernels.
Training Details
- Base model: Qwen/Qwen3.6-35B-A3B
- Fine-tuning method: SFT (full-parameter)
- Training framework: MindSpeed-MM, a distributed training acceleration framework for multimodal large models on Ascend NPUs
- Training data: TBD
Limitations
- This model is specialized for NPU kernel code generation; general conversational or general-purpose coding capabilities may be weaker than the base model.
- Generated kernel code should be compiled and numerically validated on the target hardware before production use. The model does not guarantee correctness in all scenarios.
- By using this model, you agree to comply with the Apache-2.0 license and the license terms of the base model Qwen3.6-35B-A3B.
Citation
@misc{wiki_model,
title = {wiki_model: SFT for NPU Kernel Generation},
author = {<your-name-or-org>},
year = {2026},
url = {https://huggingface.co/<your-namespace>/wiki_model}
}