---
base_model: WeiboAI/VibeThinker-3B
library_name: transformers
tags:
- gguf
- llama.cpp
- ollama
- tool-calling
- qwen2
- vibethinker
pipeline_tag: text-generation
---

# markdown-1

VibeThinker-3B fine-tuned (LoRA, merged) for **tool calling + long agent traces**.

This repo contains the **merged fp16 weights** plus ready-to-run **GGUF** quants for llama.cpp / Ollama / LM Studio.

| File | Size | Use |
|------|------|-----|
| `markdown-1-Q4_K_M.gguf` | ~1.9 GB | smaller / faster, great default |
| `markdown-1-Q8_0.gguf` | ~3.3 GB | higher fidelity |
| `model-*.safetensors` | ~6.2 GB | merged fp16 (vLLM / transformers) |

LoRA adapter only: [`notshekhar/vibethinker-finetuned-tool`](https://huggingface.co/notshekhar/vibethinker-finetuned-tool).

## Run with llama.cpp

```bash
llama-cli -hf notshekhar/markdown-1:Q4_K_M -p "Hello"
# or local:
llama-cli -m markdown-1-Q4_K_M.gguf -p "Hello"
```

## Run with Ollama

```bash
# Modelfile
printf 'FROM ./markdown-1-Q4_K_M.gguf\n' > Modelfile
ollama create markdown-1 -f Modelfile
ollama run markdown-1
```

Base reasoning model uses `<think>` traces and ChatML (`<|im_start|>`) with tool-calling via
`<tool_call>` / `<tool_response>` blocks (see `chat_template.jinja`).