---
base_model: unsloth/Qwen3-8B-Base-unsloth-bnb-4bit
tags:
- transformers
- qwen3
- Unsloth
- code
- agent
- Fine-tune
license: apache-2.0
language:
- en
datasets:
- TeichAI/MiniMax-M2.1-Code-SFT
- TeichAI/MiniMax-M2.1-8800x
- TeichAI/convo-v1
- AlicanKiraz0/Agentic-Chain-of-Thought-Coding-SFT-Dataset-v1.1
- TeichAI/claude-4.5-opus-high-reasoning-250x
pipeline_tag: text-generation
---
# LocalCodeViber

**LocalCodeViber** is a local-first agentic coding model built on [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B), fine-tuned for tool-calling, multi-step code generation, and autonomous error recovery. Designed to run entirely on consumer hardware — no API, no cloud, no cost per token.

This is the SFT foundation model. Reinforcement learning is ongoing.

---

## What it does

LocalCodeViber was trained to operate as a coding agent — not just generate code, but use tools to read files, write files, run commands, search the web, and recover from failures just like a real developer would.

It can:

- Read and edit files in a workspace
- Write complete, working code from a single prompt
- Execute shell commands and interpret the output
- Recover from failed tool calls without giving up
- Create pull requests on GitHub repositories
- Think through problems step by step using native `<think>` tags before acting

---

## Model Details

| | |
|---|---|
| **Base Model** | Qwen3-8B-Base |
| **Architecture** | Qwen3 transformer, 36 layers |

## Training Data

LocalCodeViber was trained on a curated mix of 14,837 examples across 5 datasets:

| Dataset | Examples | Focus |
|---|---|---|
| [TeichAI/convo-v1](https://huggingface.co/datasets/TeichAI/convo-v1) | 777 | Conversational format, instruction following |
| [AlicanKiraz0/Agentic-Chain-of-Thought-Coding-SFT-Dataset-v1.1](https://huggingface.co/datasets/AlicanKiraz0/Agentic-Chain-of-Thought-Coding-SFT-Dataset-v1.1) | ~3,700 | Agentic reasoning and tool use |
| [TeichAI/MiniMax-M2.1-Code-SFT](https://huggingface.co/datasets/TeichAI/MiniMax-M2.1-Code-SFT) | ~1,300 | Agentic Code generation |
| [TeichAI/MiniMax-M2.1-8800x](https://huggingface.co/datasets/TeichAI/MiniMax-M2.1-8800x) | 8,800 | Diverse coding tasks |
| [TeichAI/claude-4.5-opus-high-reasoning-250x](https://huggingface.co/datasets/TeichAI/claude-4.5-opus-high-reasoning-250x) | 250 | High-quality reasoning traces |

The dataset mix emphasises real agentic tool-use patterns including failed tool calls that are identified, diagnosed, and corrected — giving the model genuine error recovery capability rather than just pattern matching on success cases.

---

## Tools

LocalCodeViber understands the following tool schema out of the box:

```json
["read_file", "write_file", "edit_file", "list_directory", "search_code", "run_command", "web_search"]
```

These match the tools in the training data. Pass them via the standard OpenAI tool calling API.

---

## Usage

### LM Studio (Recommended)

1. Download the GGUF version: [Bob-the-Koala/LocalCodeViber-GGUF](https://huggingface.co/Bob-the-Koala/LocalCodeViber-GGUF)
2. Load in LM Studio and break free from API costs!

### Ollama

```bash
ollama run hf.co/Bob-the-Koala/LocalCodeViber-GGUF:Q4_K_M
```

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Bob-the-Koala/LocalCodeViber",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Bob-the-Koala/LocalCodeViber")
```

---

## GGUF Versions

Available in [Bob-the-Koala/LocalCodeViber-GGUF](https://huggingface.co/Bob-the-Koala/LocalCodeViber-GGUF):

| Quantization | Size | Use case |
|---|---|---|
| `Q4_K_M` | ~4.8 GB | Everyday use, best balance |

---

## System Prompt

For best results, use this system prompt:

```
You are a helpful coding assistant with access to file operations and code analysis tools.
Complete the user's task thoroughly and efficiently.
When given a coding task, create working code files in the workspace.
```

---

## Limitations

- Base model started from bnb-4bit weights — quality ceiling is below a full precision 8B model
- SFT only — reinforcement learning is in progress and will significantly improve reasoning quality
- Not suitable for tasks requiring knowledge past Qwen3's training cutoff

---

## Roadmap

- [ ] **LocalCodeViber-RL** — reinforcement learning on top of this SFT base, optimising for code correctness and task completion
- [ ] **LocalCodeViber-Claw** — fine-tuned specifically for [OpenClaw](https://github.com/openclaw/openclaw) skill schemas, channel routing, extra safety, and memory system
- [ ] **LocalCodeViber-14B** — same training recipe on Qwen3-14B for substantially higher capability

---

## Acknowledgements

LocalCodeViber was trained using [Unsloth](https://github.com/unslothai/unsloth) and would not exist without the datasets provided by [TeichAI](https://huggingface.co/TeichAI) and [AlicanKiraz0](https://huggingface.co/AlicanKiraz0).

---

## License

This model is released under the Apache 2.0 license

---

*Built by [Bob-the-Koala](https://huggingface.co/Bob-the-Koala)*
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)