CherryPit-1.1-2B-GGUF

CherryPit-1.1-2B is the second release of CherryPit from DRUPEs (Dense Representation & Unified Parameter Encoding Systems), built on MiniCPM5-2B. It reasons much better inside a fixed thinking budget: with the same 8K thinking tokens it solves far more competition-math and coding problems than CherryPit-1.0, and it can now build React web apps from tests. Same size, architecture and speed as 1.0.

Results vs CherryPit-1.0

Both models were run on the same harness, items, prompts and seeds (bf16 weights in vLLM). "Thinking, 8K" means the model may think for up to 8,192 tokens before it answers; "–" = not measured for that model.

Pooled over the 163 competition problems: 51.7. It also gets there with fewer tokens: on LiveCodeBench with thinking it uses 6,176 tokens per problem on average vs 7,454 for CherryPit-1.0.

Quantization

The Q4_K_M file was unpacked back to bf16 and re-run on the same harness, so this is what 4-bit costs. Thinking-mode coding is the most sensitive; if you have the memory, Q5_K_M or larger keeps more of it.

What's new in 1.1

• Budget-aware reasoning: reinforcement learning under a fixed thinking budget, so the model reaches an answer inside 8K tokens instead of running out of room.
• Ladder training on hard problems: the model starts from hints taken from its own long correct solutions, and the hints are removed step by step until it solves the problem on its own (math and code).
• Knowing when to stop: once it has found the answer it commits to it instead of thinking on.
• Web apps: React components written to pass a given set of tests (from the 1.1 WebDev preview).

Model Details

• Developed by: DRUPEs
• Model type: Causal language model
• Parameters: ~2B
• Base model: MiniCPM5-2B
• Previous version: CherryPit-1.0-2B-GGUF
• 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: 1.1

How to Use

Which file should I pick? Start with Q4_K_M (1.56 GB). If you have the memory, Q6_K or Q8_0 are close to lossless. Avoid the Q2 files.

Ollama

ollama run hf.co/DRUPEs/CherryPit-1.1-2B-GGUF:Q4_K_M

LM Studio

Search for DRUPEs/CherryPit-1.1-2B-GGUF in the Discover tab and pick a quant.

llama.cpp

# chat in the terminal
llama-cli -hf DRUPEs/CherryPit-1.1-2B-GGUF:Q4_K_M

# OpenAI-compatible server on http://localhost:8080
llama-server -hf DRUPEs/CherryPit-1.1-2B-GGUF:Q4_K_M

Recommended sampling: temperature 1.0, top_p 0.95.

Thinking on or off

The chat template starts in thinking mode, which is best for hard math and code; give it 8K or more new tokens. For short, direct answers turn thinking off: in llama.cpp add --chat-template-kwargs '{"enable_thinking":false}'; in other apps use the "thinking" or "reasoning" toggle. With no system prompt the model introduces itself as CherryPit.

Files

All quants use an importance matrix computed at 4K context on CherryPit's own long reasoning traces plus general chat and code, which keeps thinking-mode quality far better than a short-answer calibration (Q4_K_M: 41.1 vs 36.0 on LiveCodeBench with thinking). _L = token embedding and output layer kept at Q8_0. The F16 file is full precision.

Limitations

• It is a 2B model: it can be confidently wrong, so check answers that matter.
• Instruction following (IFEval) and HumanEval+ are slightly below 1.0, see the table.
• Tested in English only.