---
license: apache-2.0
base_model: Jackrong/Qwopus3.5-4B-Coder
datasets:
- Glint-Research/Fable-5-traces
language:
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- qwen3_5
- fable5
- reasoning
- chain-of-thought
- agent
- tool-use
- function-calling
- coder
- coding
- debugging
- conversational
- unsloth
- lora
- sft
- safetensors
---

<div align="center">

# 💻 Qwopus3.5-4B-Coder-Fable5-v1

### Fable-5 trace continuation of Qwopus3.5-4B-Coder

<p>
  <b>Agentic coding</b> ·
  <b>tool use</b> ·
  <b>debugging</b> ·
  <b>local inference</b>
</p>


</div>

---

## Overview

**Qwopus3.5-4B-Coder-Fable5-v1** is a Fable-5 trace continuation of [`Jackrong/Qwopus3.5-4B-Coder`](https://huggingface.co/Jackrong/Qwopus3.5-4B-Coder).

The base model, Qwopus3.5-4B-Coder, is a compact Qwen3.5-based coding model trained for reasoning, tool use, function calling, coding workflows, and agent-style behavior.

This release continues that model on [`Glint-Research/Fable-5-traces`](https://huggingface.co/datasets/Glint-Research/Fable-5-traces), a dataset of Claude Fable 5 local coding-agent traces. The dataset is heavily oriented around tool-use trajectories, repository work, local command context, code editing, debugging loops, and `<think>`-style reasoning completions.

The result is a small local coding-agent model intended for:

| Area | Description |
|---|---|
| Tool-use workflows | Bash, Read, Write, Edit, repo inspection, and action traces. |
| Debugging | Failing tests, stack traces, root-cause analysis, and patch planning. |
| Trace-style reasoning | Long-form planning and `<think>` style reasoning traces. |
| Local agents | Hermes-style, Claude-Code-style, OpenCode-style, and LM Studio workflows. |



## About the Fable-5 Traces

[`Glint-Research/Fable-5-traces`](https://huggingface.co/datasets/Glint-Research/Fable-5-traces) contains Claude Fable 5 coding traces.

The dataset includes fields such as:

```text
uid
source_file
session
model
context
cot
output_type
output
completion
origin
```

The examples are not simple chat pairs. They are multi-step agent trajectories with local development context, reasoning traces, and tool-use outputs.

Common patterns in the dataset include:

- user coding requests
- local-command caveats
- repository inspection
- Bash command usage
- file reads
- file writes
- edits
- debugging passes
- playtesting / validation loops
- `<think>...</think>` reasoning traces
- tool-use completions

A large portion of the dataset is `tool_use` style data, which makes it especially relevant for local coding agents and developer automation.

## Capabilities

### Agentic coding

Designed for coding-agent loops where the model must inspect a repo, plan work, call tools, edit files, and validate changes.

### Tool-use style outputs

Works well with prompts that expose structured tools such as:

```text
Bash
Read
Write
Edit
Search
Grep
```

### Debugging and repair

Useful for:

- finding likely failing files
- explaining stack traces
- planning test commands
- proposing minimal patches
- iterating after errors

### Local-first deployment

The release includes Transformers, GGUF, MLX, and MLX 4-bit formats so it can run in Python, llama.cpp, LM Studio, and Apple Silicon workflows.

## Quick Start

```python
import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM

model_id = "shuhulx/Qwopus3.5-4B-Coder-Fable5-v1"

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
)

model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "Inspect this repo and write a Bash/Read/Edit style plan for debugging failing tests."
            }
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=768,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
)

print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

## Available Releases

| Release | Repo | Best for |
|---|---|---|
| Transformers / Safetensors | [`shuhulx/Qwopus3.5-4B-Coder-Fable5-v1`](https://huggingface.co/shuhulx/Qwopus3.5-4B-Coder-Fable5-v1) | Python, Transformers, custom inference. |
| GGUF | [`shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-GGUF`](https://huggingface.co/shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-GGUF) | llama.cpp, LM Studio, local CPU/GPU inference. |
| MLX | [`shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-MLX`](https://huggingface.co/shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-MLX) | Apple Silicon full MLX inference. |
| MLX 4-bit | [`shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-MLX-4bit`](https://huggingface.co/shuhulx/Qwopus3.5-4B-Coder-Fable5-v1-MLX-4bit) | Apple Silicon low-memory inference. |


## Credits

Built on:

- [`Jackrong/Qwopus3.5-4B-Coder`](https://huggingface.co/Jackrong/Qwopus3.5-4B-Coder) by Jackrong
- [`Glint-Research/Fable-5-traces`](https://huggingface.co/datasets/Glint-Research/Fable-5-traces) by Glint-Research
- Qwen / Qwen3.5 model family
- Unsloth
- Hugging Face
- llama.cpp
- mlx-lm