CherryPit-1.0-2B-GGUF
CherryPit-1.0-2B is a ~2B-parameter language model from DRUPEs (Dense Representation & Unified Parameter Encoding Systems), built on MiniCPM5-2B. It's a drop-in upgrade with overall improvements: a comparatively large gain in coding, small gains in math and reasoning, and the same size, architecture and speed as the base model.
Results vs MiniCPM5-2B
Both models were run on the same harness, items, prompts and seeds. Thinking mode off.
| Benchmark | Setting | MiniCPM5-2B | CherryPit-1.0-2B | Δ |
|---|---|---|---|---|
| Coding | ||||
| LiveCodeBench v6 (2025-01..04) | pass@1, avg of 8, T=1.0 | 30.4 | 32.4 | +1.9 |
| — easy / medium / hard | 81.7 / 27.2 / 5.0 | 84.3 / 28.8 / 6.7 | +2.6 / +1.6 / +1.7 | |
| MBPP+ (EvalPlus) | greedy | 61.9 | 65.1 | +3.2 |
| HumanEval | greedy | 89.6 | 90.9 | +1.2 |
| HumanEval+ | greedy | 85.4 | 86.0 | +0.6 |
| General | ||||
| MATH-500 | greedy | 79.4 | 79.8 | +0.4 |
| BBH | 0-shot CoT | 77.8 | 78.4 | +0.6 |
| MMLU-Pro | answer-letter scoring | 35.0 | 34.8 | −0.2 |
| HotpotQA (distractor) | F1 | 55.1 | 54.7 | −0.4 |
| HellaSwag | acc_norm | 50.5 | 49.5 | −1.0 |
Answer length is unchanged (≈2,100 tokens per LiveCodeBench answer for both models).
Model Details
- Developed by: DRUPEs
- Model type: Causal language model
- Parameters: ~2B
- Base model: MiniCPM5-2B
- Architecture: Standard Llama (381 tensors)
- Context length: 128K tokens (131,072)
- Tool calling: Supported (via the built-in chat template)
- Precision: bf16
- Language(s): English
- License: Apache 2.0
- Version: 1.0
- Full-precision weights: DRUPEs/CherryPit-1.0-2B (bf16 safetensors)
How to Use
Which file should I pick? Start with Q4_K_M (1.45 GB). If you have the memory, Q6_K is effectively lossless. Avoid the Q2 files.
Ollama
ollama run hf.co/DRUPEs/CherryPit-1.0-2B-GGUF:Q4_K_M
LM Studio
Search for DRUPEs/CherryPit-1.0-2B-GGUF in the Discover tab and pick a quant.
llama.cpp
# chat in the terminal
llama-cli -hf DRUPEs/CherryPit-1.0-2B-GGUF:Q4_K_M
# OpenAI-compatible server on http://localhost:8080
llama-server -hf DRUPEs/CherryPit-1.0-2B-GGUF:Q4_K_M
Recommended sampling: temperature 1.0, top_p 0.95 (the defaults stored in the GGUF).
Files
GGUF files work with llama.cpp, LM Studio, Ollama and other GGUF runtimes. Quality loss is measured against the full-precision model; lower is better.
| File | Size | Quality loss vs full precision |
|---|---|---|
CherryPit-1.0-2B-Q8_0.gguf |
2.50 GB | +0.3% (lossless) |
CherryPit-1.0-2B-Q6_K_L.gguf |
2.05 GB | −0.2% (lossless) |
CherryPit-1.0-2B-Q6_K.gguf |
1.93 GB | −0.1% (lossless) |
CherryPit-1.0-2B-Q5_K_M.gguf |
1.68 GB | +1.1% |
CherryPit-1.0-2B-Q4_K_L.gguf |
1.64 GB | +1.9% |
CherryPit-1.0-2B-Q4_K_M.gguf |
1.45 GB | +2.0%, recommended default |
CherryPit-1.0-2B-Q4_K_S.gguf |
1.40 GB | +2.4% |
CherryPit-1.0-2B-IQ4_XS.gguf |
1.33 GB | +3.4% |
CherryPit-1.0-2B-Q3_K_L.gguf |
1.28 GB | +7.8% |
CherryPit-1.0-2B-Q3_K_M.gguf |
1.20 GB | +8.4% |
CherryPit-1.0-2B-IQ3_M.gguf |
1.14 GB | +10.6% |
CherryPit-1.0-2B-IQ3_XS.gguf |
1.07 GB | +16.8% |
CherryPit-1.0-2B-Q2_K_L.gguf |
1.21 GB | +51% (skip) |
CherryPit-1.0-2B-IQ2_M.gguf |
0.90 GB | +66% (skip) |
Training
🔒 The full method will be released once the model reaches 30,000 downloads or on October 31, 2026, whichever comes first.
Limitations
- CherryPit is a 2B model and shares the limits of its base: it can make factual errors and write buggy code, so review its output before relying on it.
- The largest gains are in coding; improvements in other areas are smaller.
- Evaluations were run in English with thinking mode off.
Citation
@misc{drupes2026cherrypit,
title = {CherryPit-1.0-2B},
author = {DRUPEs},
year = {2026},
url = {https://huggingface.co/DRUPEs/CherryPit-1.0-2B-GGUF}
}