Atomic-Germ/Qwen3.8-Distilled-4B-NPU2

PASSES EVERY FLM-TEST SUITE

IF YOU USE COMMUNITY QWEN MODELS DO NOT UPGRADE TO FLM v1.0.2+

Qwen3.8 4B Distilled — Q4NX

FastFlowLM Q4NX conversion of `Ma7ee7/Qwen3.8_4B_Distilled` for AMD XDNA NPU inference.

This repository contains a quantized Q4NX port of the model, compiled for the FastFlowLM (FLM) runtime. It is not a GGUF file.

Qwen3.8 4B Distilled is a 4-billion-parameter reasoning model created by distilling outputs from Qwen3.8-Max into the smaller Qwen3-4B-Thinking-2507 student model.

The full model was fine-tuned on `r0b0tlab/qwen3.8-max-distillation-50k`, a dataset of responses and reasoning traces generated by qwen3.8-max-preview.

What is Q4NX?

Q4NX is FastFlowLM's native packed-quantization format - a rearranged Q4_1 layout tuned for the NPU matrix engine's tile sizes and memory access patterns. It is not a GGUF file and it does not run on llama.cpp or Ollama; it is meant exclusively for the FastFlowLM engine on AMD Ryzen AI NPUs.

Requirements

• —FastFlowLM >= 0.9.45 (flm CLI)
• —AMD Ryzen AI processor with XDNA2 (NPU2) - Strix Point / Ryzen AI 300 series or later
• —Linux with the XRT NPU stack installed
• —~16 GB of unified system memory (Q4NX weights + activations + KV cache)

FLM Bench

Tested on an AMD Ryzen AI 340 Framework 13 laptop.

Files

Install and run

This repository works with flm-add, a small installer that copies the model into the FastFlowLM user directory and registers the tag. It never modifies the system FastFlowLM install.

pip install flm-add or uv tool install flm-add

uv tool install flm-add
flm-add Atomic-Germ/Qwen3.8-Distilled-4B-NPU2 --family qwen3.5
FLM_CONFIG_PATH="$HOME/.config/flm/model_list.json" FLM_XCLBIN_PATH="$HOME/.config/flm" flm run qwen3.8-distilled:4b

Kernels

FastFlowLM's NPU kernels (xclbins) are closed source and are not shipped in this repository. flm-add links the kernels of the official `qwen3-tk:4b` model (Qwen3-4B-Thinking-2507-NPU2), because this model shares the same engine family (qwen3) and architecture.

Model Lineage and Naming

The name Qwen3.8 4B Distilled describes the model's distillation lineage:

• —Teacher model: qwen3.8-max-preview
• —Student/base model: Qwen/Qwen3-4B-Thinking-2507
• —Full-precision fine-tune: Ma7ee7/Qwen3.8_4B_Distilled
• —Resulting model size: Approximately 4 billion parameters
• —Distillation dataset: r0b0tlab/qwen3.8-max-distillation-50k
• —Distribution format: Q4NX

This is a Qwen3-architecture student model distilled from Qwen3.8-Max-generated outputs.

The repository does not claim that the underlying architecture or original weights are from Qwen3.8-Max. Qwen3.8-Max is the teacher whose generated responses and reasoning traces were used as training targets for the 4B student.

This repository contains quantized conversions of that independently fine-tuned student model. It is not an official Qwen or Alibaba release.

Links

• —GGUF repository: Ma7ee7/Qwen3.8_4B_Distilled_GGUF
• —Full model: Ma7ee7/Qwen3.8_4B_Distilled
• —Base model: Qwen/Qwen3-4B-Thinking-2507
• —Training dataset: r0b0tlab/qwen3.8-max-distillation-50k

Model Details

What “Distilled” Means Here

This model uses sequence-level knowledge distillation.

The smaller student was trained on complete responses produced by the larger teacher. This transfers parts of the teacher's behavior, reasoning patterns, solution structure, and response style without copying the teacher's architecture or weights.

Therefore:

• —The architecture and original student weights come from Qwen3-4B-Thinking-2507.
• —The distillation targets come from Qwen3.8-Max-generated outputs.
• —The resulting checkpoint remains a 4B Qwen3 model.
• —These Q4NX file is a quantized versions of that 4B distilled model.
• —The model is not expected to reproduce the full capabilities of Qwen3.8-Max.

Thinking Output

The model inherits a thinking-oriented chat format from Qwen3-4B-Thinking-2507. Depending on the inference application and reasoning parser, visible reasoning may be displayed in a form similar to:

[FLM]  Model RAW Output: 
<think>
The user is asking a straightforward factual question: "What is the capital of France?"

Since no specific tools or functions are available for answering such questions, I'll simply provide the answer directly.
</think>

The capital of France is **Paris**.

Some applications may hide the thinking section, place it in a collapsible panel, or display it separately from the final answer.

Distillation Dataset

The full model was trained on:

`r0b0tlab/qwen3.8-max-distillation-50k`

The dataset contains teacher-generated examples across areas including:

• —Mathematics
• —Programming
• —General reasoning
• —Scientific reasoning
• —Instruction following
• —Limited tool use

Teacher responses were generated by qwen3.8-max-preview. Visible <think>...</think> reasoning traces were retained when present in the dataset.

Intended Uses

This model is intended for local experimentation with:

• —Mathematical reasoning
• —Programming and code generation
• —Logical reasoning
• —Scientific question answering
• —General instruction following
• —Long-form problem solving
• —Local conversational assistants
• —Research into teacher-to-student distillation
• —FastFlowLM inference

Limitations

• —This is a quantized 4B student model and does not contain the complete capabilities of Qwen3.8-Max.
• —Quantization may reduce quality relative to the full-precision model.
• —Lower-bit quantizations may degrade more noticeably on difficult reasoning or programming tasks.
• —Distillation transfers patterns from teacher-generated outputs; it does not copy the teacher's architecture or weights.
• —Teacher-generated answers may contain factual, mathematical, or programming errors.
• —Visible reasoning traces should not automatically be assumed to be correct.
• —The model may hallucinate information or produce confidently incorrect answers.
• —The training mixture is predominantly English.
• —The training dataset includes examples originating from common evaluation benchmarks.
• —Results on overlapping benchmarks may be contaminated and should not be treated as clean evaluations without additional controls.
• —Outputs should be reviewed before use in high-stakes, medical, financial, legal, or security-sensitive settings.

License and Training-Data Notice

This model repository is published under the Apache License 2.0.

This license does not override the licenses, attribution requirements, or usage terms associated with:

• —The full-precision fine-tuned model
• —The Qwen3 base model
• —The Qwen3.8-Max teacher provider
• —The distillation dataset
• —The original datasets from which prompts were sourced

Users are responsible for reviewing the relevant upstream model, dataset, and provider terms before redistribution or commercial use.

Acknowledgements

This model builds upon work from:

• —The Qwen team for Qwen3-4B-Thinking-2507
• —r0b0tlab for the Qwen3.8-Max Distillation 50K dataset
• —llama.cpp and the GGUF ecosystem
• —AMD's FastFlowLM team and NPU
• —Hugging Face

Citation

Distillation Dataset

@misc{r0b0tlab2026qwen38distillation50k,
  title        = {Qwen3.8-Max Distillation 50K},
  author       = {r0b0tlab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-distillation-50k}}
}

Disclaimer

Qwen3.8 4B Distilled GGUF is an independent community release by Ma7ee7.

It is not produced, endorsed, or officially released by the Qwen team, Alibaba, or Alibaba Cloud.

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report},
      author={Qwen Team},
      year={2025},
      eprint={2505.09388},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.09388},
}