---

license: apache-2.0

base_model:

• openbmb/MiniCPM5-2B

pipeline_tag: text-generation

tags:

• text-generation-inference
• llama-cpp
• minicpm
• minicpm5
• llama
• text-generation
• long-context
• tool-calling
• on-device
• edge-ai

datasets:

• openbmb/Ultra-FineWeb
• openbmb/UltraX-Preview
• openbmb/Ultra-FineWeb-L3
• openbmb/UltraData-Math
• openbmb/UltraData-Code
• openbmb/UltraData-SFT-2605
• openbmb/UltraData-SFT-Agent-2609
• openbmb/UltraData-RL-2609

language:

• en

library_name: transformers

---

MiniCPM5-2B-GGUF

> MiniCPM5-2B is the second model in OpenBMB's MiniCPM5 series, a dense 2.5-billion-parameter (2.52B) causal language model built on the standard LlamaForCausalLM architecture (42 layers, GQA with 16 Q heads / 2 KV heads, 131,072-token context), designed for local assistants, coding agents, tool-use workflows, and on-device deployment. It's trained through a full UltraData Tiered Data Management pipeline — base training, mid-training on curated corpora like Ultra-FineWeb, UltraX, and UltraData-Code/Math, then post-training via 400B tokens of deep-thinking SFT (including a 500K-sample agentic dataset), followed by domain-specialized RL teachers (math, code, agentic, writing) and On-Policy Distillation (OPD) that merges all 16 expert models back into one release checkpoint using full-vocabulary reverse-KL advantage estimation, together delivering an average +10.96 point gain in reasoning/general capabilities and +6.96 in agentic capabilities over the SFT-only checkpoint. Within its comparison set of 2B-class open models, MiniCPM5-2B achieves an average benchmark score of 53.9 — exceeding even several 4B-class models like Qwen3.5-4B (51.1) — with particularly strong results in code reasoning (69.1 on LiveCodeBench v6), math (86.5 on AIME 2025/2026), long context (68.1 on NoLiMa), tool use (97.1 on τ²-Bench Telecom), and coding-agent tasks (46.4 on SWE-bench Verified). It's servable via vLLM, SGLang (recommended for XML-style tool calling with the native minicpm5 parser), Transformers, llama.cpp/GGUF, Ollama, LM Studio, MLX, or across nine AI chip architectures via the FlagOS ecosystem, with a companion MiniCPM5-2B-DSpark draft model available for speculative decoding, and is released under the Apache-2.0 license.

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| MiniCPM5-2B.BF16.gguf | BF16 | 5.04 GB | Download |

| MiniCPM5-2B.F16.gguf | F16 | 5.04 GB | Download |

| MiniCPM5-2B.F32.gguf | F32 | 10.1 GB | Download |

| MiniCPM5-2B.Q3_K_L.gguf | Q3_K_L | 1.38 GB | Download |

| MiniCPM5-2B.Q3_K_M.gguf | Q3_K_M | 1.29 GB | Download |

| MiniCPM5-2B.Q3_K_S.gguf | Q3_K_S | 1.19 GB | Download |

| MiniCPM5-2B.Q4_0.gguf | Q4_0 | 1.49 GB | Download |

| MiniCPM5-2B.Q4_K_M.gguf | Q4_K_M | 1.56 GB | Download |

| MiniCPM5-2B.Q4_K_S.gguf | Q4_K_S | 1.5 GB | Download |

| MiniCPM5-2B.Q5_0.gguf | Q5_0 | 1.77 GB | Download |

| MiniCPM5-2B.Q5_K_M.gguf | Q5_K_M | 1.81 GB | Download |

| MiniCPM5-2B.Q5_K_S.gguf | Q5_K_S | 1.77 GB | Download |

| MiniCPM5-2B.Q6_K.gguf | Q6_K | 2.07 GB | Download |

| MiniCPM5-2B.Q8_0.gguf | Q8_0 | 2.68 GB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp