CherryPit-1.2-2B

CherryPit-1.2-2B

A 2B coding and chat model built on MiniCPM5-2B that is much better at web apps needing several features in one component. For developers who want a small local model for code.

Made by DRUPEs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: CherryPit-1.2-2B comes from DRUPEs and is the newest of three CherryPit releases. It is built on MiniCPM5-2B. The big gain over 1.1 is combining features. Pages that must pack two to four features into a single component get through the tests much more often. It also stops adding its own name to ordinary answers.

Who it suits: Developers who want a small model for code, web components and everyday tasks that they can run themselves.

Industry | Who uses it | How they use it | Example
Software | Web developers | Generate React components with several features | A component with search and a loading list
Education | Students | Practice coding with a small local model | Ask for a Python function
Research | Small-model researchers | Study gains from composition training | Compare 1.1 and 1.2 on WebApp1K
Platform | Tooling teams | Serve a small model with vLLM | A local coding endpoint

You can skip it if you need reliable answers from a model this small. The card warns it can sound sure and still be wrong, and that four features in one component pass only 28 percent of the time.

Where to use it:

• Python, through Hugging Face Transformers
• vLLM, as an OpenAI-style server
• llama.cpp, Ollama or LM Studio, through the separate GGUF repository
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Web component code | Developers | React code that combines several features
Coding help | Students | Python and general code answers
Short answers | Anyone | Direct replies when thinking is turned off

How to set it up:

• Python and pip on your computer
• Transformers or vLLM, installed
• A GPU or enough memory for about 5 GB of BF16 weights, or about 1.3 GB at 4-bit by estimate

1. Install the Hugging Face hub package.
2. Download the model files with the hf command.
3. Run the Python example or serve the model with vLLM.

pip install -U huggingface_hub
hf download DRUPEs/CherryPit-1.2-2B

How to use it: Load the model in Transformers, apply the chat template and generate, or serve it with vLLM. The card recommends temperature 1.0 and top_p 0.95. The template starts in thinking mode. Pass enable_thinking as False for short, direct answers.

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("DRUPEs/CherryPit-1.2-2B")
model = AutoModelForCausalLM.from_pretrained("DRUPEs/CherryPit-1.2-2B", torch_dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "Build a React component with a search box and a list that loads on open."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=8192, do_sample=True, temperature=1.0, top_p=0.95)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

vllm serve DRUPEs/CherryPit-1.2-2B --max-model-len 32768

Example: You ask for a React component that combines a search field with a list that fills when the page opens. It returns the component code, and in thinking mode it reasons first.

For newcomers and experts: If you are new, run the Python example and check any answer that matters. This is a 2B model and can sound sure while being wrong.

If you know the field, 1.2 adds reinforcement-learning training on made-up web apps with several features and on code tasks with several steps. Rounds of training on its own passing answers follow. The card reports WebApp1K Trio at 44.9 against 16.8 for 1.1, tested on a benchmark it never trained on.

MiniCPM5-2B base model
Composition training on multi-feature apps
Self-taught rounds on passing answers
CherryPit-1.2-2B

Title: WebApp1K Trio pass rate, three features in one component
CherryPit-1.1: 16.8
CherryPit-1.2: 44.9

Good to know:

• License: Apache 2.0
• Cost to run: your own hardware, with no paid service named
• Data it reaches: the files you download and the prompts you give it
• Limits: tested in English only, and long math with 32K thinking is about level with 1.1
• Identity: with thinking off, about 1 in 20 identity questions still names its base model

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 2.5B
• Weights: about 5 GB in BF16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 451
• Likes: 6
• License: apache-2.0
• Languages: en
• Based on: openbmb/MiniCPM5-2B
• Quantized versions: 4
• Tags: transformers, safetensors, llama, text-generation, code, coding, web-development, math, reasoning, tool-calling, on-device, small-language-model

Links

• Model: https://huggingface.co/DRUPEs/CherryPit-1.2-2B
• Files: https://huggingface.co/DRUPEs/CherryPit-1.2-2B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download DRUPEs/CherryPit-1.2-2B

Then follow the usage in its model card. Nothing in this kit runs by itself.