Instructions to use DRUPEs/CherryPit-1.2-2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use DRUPEs/CherryPit-1.2-2B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="DRUPEs/CherryPit-1.2-2B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("DRUPEs/CherryPit-1.2-2B")
model = AutoModelForCausalLM.from_pretrained("DRUPEs/CherryPit-1.2-2B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use DRUPEs/CherryPit-1.2-2B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "DRUPEs/CherryPit-1.2-2B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DRUPEs/CherryPit-1.2-2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/DRUPEs/CherryPit-1.2-2B

SGLang

How to use DRUPEs/CherryPit-1.2-2B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "DRUPEs/CherryPit-1.2-2B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DRUPEs/CherryPit-1.2-2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "DRUPEs/CherryPit-1.2-2B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DRUPEs/CherryPit-1.2-2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use DRUPEs/CherryPit-1.2-2B with Docker Model Runner:
```
docker model run hf.co/DRUPEs/CherryPit-1.2-2B
```

CherryPit-1.2-2B

CherryPit-1.2-2B is the third release of CherryPit from DRUPEs (Dense Representation & Unified Parameter Encoding Systems), built on MiniCPM5-2B. It is much better at combining things: web apps that need two, three or four features in one component now pass 1.5x to 4x as often as with CherryPit-1.1, while single features, coding, math and instruction following all went up a little. It also stops dropping its own name into ordinary answers. Same size, architecture and speed as 1.0 and 1.1.

Results vs CherryPit-1.1

Both models were run on the same harness, items, prompts and seeds (bf16 weights in vLLM). We fixed a harness bug (text-path evals were fed two start-of-text tokens), so 1.1's numbers here are re-measured and differ from its card.

Benchmark	Setting	CherryPit-1.1	CherryPit-1.2	Δ
Web development
WebApp1K-React, 1 feature	no thinking, 1,000 tasks	81.9	82.7	+0.8
WebApp1K Duo, 2 features in one component	no thinking, 1,000 tasks	31.6	48.8	+17.2
WebApp1K Trio, 3 features	no thinking, 1,000 tasks	16.8	44.9	+28.1
WebApp1K Quartet, 4 features	no thinking, 1,000 tasks	6.6	28.3	+21.7
Coding
LiveCodeBench v6	no thinking, avg of 4	36.6	38.3	+1.7
LiveCodeBench v6	thinking, 8K	45.7	47.4	+1.7
HumanEval+	no thinking, greedy	82.9	84.1	+1.2
MBPP+	no thinking, greedy	67.2	68.0	+0.8
Math
MATH-500	no thinking, greedy	87.8	90.6	+2.8
AIME 2025 + 2026	thinking, 32K, 2 samples each	70.0	67.5	−2.5 (not significant)
Instruction following
IFEval	strict, prompt level	87.4	88.7	+1.3

The web-app training used only our own generated tasks, never WebApp1K, so the Duo / Trio / Quartet gains are on a benchmark the model has not seen.
No benchmark got significantly worse: every row was checked per item (paired sign tests) against 1.0 and 1.1.

What's new in 1.2

Composition training: reinforcement learning on generated multi-feature web apps and multi-step code pipelines.
Self-taught rounds: the model trained on its own answers that passed the tests on combined features, mixed with its own correct single-feature and Python answers so it did not forget anything.
A better default prompt: with no system prompt, the model introduces itself only when it is asked who it is.

Model Details

Developed by: DRUPEs
Model type: Causal language model
Parameters: ~2B
Base model: MiniCPM5-2B
Previous version: CherryPit-1.1-2B
Architecture: Standard Llama
Context length: 128K tokens (131,072)
Tool calling: Supported (via the built-in chat template)
Language(s): English
License: Apache 2.0
Version: 1.2

How to Use

Full-precision (bf16) weights for transformers, vLLM and further fine-tuning. For llama.cpp, Ollama or LM Studio use DRUPEs/CherryPit-1.2-2B-GGUF.

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

Recommended sampling: temperature 1.0, top_p 0.95. The template starts in thinking mode; pass enable_thinking=False to apply_chat_template (vLLM: chat_template_kwargs) for short, direct answers.

Limitations

It is a 2B model: it can be confidently wrong, so check answers that matter.
Combining features is much better but not solved: four features in one component pass 28% of the time vs 83% for one.
Long-budget math (32K-token thinking) is about level with 1.1, not better.
With thinking off, about 1 in 20 answers to "What model are you?"-type questions still gives its base model's name.
Tested in English only.