README.md

---
license: other
language:
- en
pipeline_tag: text-generation
tags:
- qwen2.5
- coder
- coding
- fastflowlm
- q4nx
- npu
base_model:
- Qwen/Qwen2.5-Coder-3B
---

# Qwen2.5-3B-Coder-Instruct - Q4NX for FastFlowLM (AMD Ryzen AI XDNA2)

Qwen2.5-Coder-3B Instruct, converted to Q4NX for FastFlowLM. Runs on the Qwen2.5 engine (3B kernels).

## 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](https://fastflowlm.com)
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
- ~8 GB of unified system memory (Q4NX weights + activations + KV cache)

## Files

| File | Purpose |
|---|---|
| model.q4nx  | Quantized Q4NX weights |
| config.json  | FastFlowLM model configuration |
| tokenizer.json  | Tokenizer |
| tokenizer_config.json  | Special tokens and chat template |
| chat_template.jinja  | Chat template (optional) |
| flm-add.py  | Installer script - registers this model with FastFlowLM |


## Install and run

This repository ships `flm-add.py`, a small installer that copies the model
into the FastFlowLM user directory and registers the tag `qwen2.5-coder:3b`. It never
modifies the system FastFlowLM install.

```bash
# one-time environment (add these to ~/.bashrc)
export FLM_CONFIG_PATH="$HOME/.config/flm/model_list.json"
export FLM_XCLBIN_PATH="$HOME/.config/flm"

git lfs install
git clone https://huggingface.co/Atomic-Germ/Qwen2.5-3B-Coder-Instruct-NPU2
cd Qwen2.5-3B-Coder-Instruct-NPU2
python3 ./flm-add.py . 
flm run qwen2.5-coder:3b
```

Run `python3 ./flm-add.py --help` for all options. Without a clone, the same
command works against the repo id directly:

```bash
python3 ./flm-add.py Atomic-Germ/Qwen2.5-3B-Coder-Instruct-NPU2 
```


## Kernels

FastFlowLM's NPU kernels (xclbins) are closed source and are not shipped in
this repository. `flm-add.py` links the kernels of the official **`qwen2.5-it:3b`**
model (`Qwen2.5-3B-Instruct-NPU2`), because this model shares the same engine family
(`qwen2`) and architecture.

## Serve (OpenAI-compatible)

```bash
flm serve qwen2.5-coder:3b --port 8080
```

```bash
curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen2.5-coder:3b","messages":[{"role":"user","content":"Hello!"}],"max_tokens":256}'
```

## Model

- Registry tag: `qwen2.5-coder:3b`
- Engine family: `qwen2`
- Kernel source: Qwen2.5-3B-Instruct-NPU2
- Context length: 32,768 tokens (from config)
- `model.q4nx` size: 2.59 GB
- Base model: [Qwen/Qwen2.5-Coder-3B](https://huggingface.co/Qwen/Qwen2.5-Coder-3B)
- License: other

## Original model card

See the upstream model card for training details, benchmarks, and upstream
usage. This repository only contains the Q4NX conversion for FastFlowLM.
- Upstream card: [Qwen/Qwen2.5-Coder-3B](https://huggingface.co/Qwen/Qwen2.5-Coder-3B)