---
title: MiniCPM5-1B-Agent
emoji: 🛠️
colorFrom: indigo
colorTo: yellow
sdk: docker
app_port: 7860
pinned: false
# Build Small Hackathon TRACK + MERIT-BADGE tags. Kebab-case best-guesses - confirm the exact strings against
# the Field Guide (https://build-small-hackathon-field-guide.hf.space/) before final submit; the intent is right.
tags:
  - backyard-ai          # TRACK: local, self-hosted AI (swap if you pick the other track)
  - well-tuned           # full fine-tune of MiniCPM5-1B, published on the Hub
  - llama-champion       # served on the llama.cpp runtime
  - off-the-grid         # runs fully local on a CPU, no cloud model APIs
---

# 🛠️ MiniCPM5-1B-Agent

**A tiny agentic coding agent that runs the whole write → run → read → debug → verify loop on a free CPU.**

![MiniCPM5-1B-Agent demo](minicpm5-1b-agent-demo.gif)

A full fine-tune of [`openbmb/MiniCPM5-1B`](https://huggingface.co/openbmb/MiniCPM5-1B) (1B params), served as a
Q8_0 GGUF on llama.cpp, no GPU. Give it a task; it reasons in `<think>`, then uses `bash` / `write` / `read` /
`edit` in a sandbox to build, run, and fix code, and renders the result (charts, images, live HTML) inline in
the chat. Multi-turn: files and history persist across messages. It is also exposed as an **MCP tool**
(`run_coding_task` at `/gradio_api/mcp/`).


## What it is

Most coding agents are 70B+ behind a cloud API. This is the opposite: a **1B** model doing the *real* agentic
loop on a **2-vCPU CPU Space**, no GPU. It writes a file, runs it, reads the output, debugs, and shows you the
artifact, the same loop a big agent runs, shrunk to something you could host in your own backyard.

## How it was built

- **Data (`train_v4`, 45,762 rows):** the proven v2 backbone (retail-filtered teacher mixes + real-usage agent
  traces) kept whole, plus ~3,538 curated additions, gated to a small served tool vocab and solution-aware
  MinHash-deduped. Bundled on the model repo under `dataset/`.
- **SFT:** full fine-tune (not LoRA, the long agentic mix needs the capacity) of the abliterated base, 1 epoch,
  24k context, fits in ~15-18 GB VRAM (direct Liger fused cross-entropy + mem-efficient SDPA).
- **DPO (on-policy):** run the SFT model over the training prompts and capture its OWN behaviour. *chosen* = a
  valid `<function>` tool call; *rejected* = its real miss (rambling in `<think>` / answering with no call).
  ~649 pairs. This rewards ACTING over stalling.
- **Serving:** Q8_0 GGUF on llama.cpp; a two-phase decode bounds the `<think>` separately from the action so the
  model acts instead of looping; produced files render inline (charts, images, sandboxed live-HTML iframes).

## Try it

- "Write a Python script that makes a bar chart of 30, 45, 25 (A, B, C), save chart.png, then run it." → writes + runs it; the PNG renders inline.
- "Write an HTML page quote.html with a button that shows a random quote each click (hard-coded, no internet)." → writes the file; renders live in a sandboxed iframe.

It is a tiny 1B on a free CPU: expect **~4 min per simple turn**, longer for multi-step tasks (the demo video
shows it working end-to-end, so it can be judged even if a live run is slow).

## Model, dataset & full reproduction

→ **[Luminia/MiniCPM5-1B-Agent-GGUF](https://huggingface.co/Luminia/MiniCPM5-1B-Agent-GGUF)** (model card =
the full data mix, SFT/DPO recipe, eval, and exact reproduce commands; v4 dataset bundled under `dataset/`).

💻 **Code on GitHub:** [Katehuuh/MiniCPM5-1B-Agent](https://github.com/Katehuuh/MiniCPM5-1B-Agent) (the Space + the full training pipeline; code reviewed with OpenAI Codex).

*Built for the Build Small Hackathon · OpenBMB + OpenAI Codex tracks.*