Model

MiniCPM-2B

OPEN

OpenBMB · MiniCPM family · released Feb 1, 2024

Pioneering ~2.4B-non-embedding edge LLM matching Mistral-7B on many benchmarks; runs on-device including smartphones

ReasoningCodingVisionFunction callingTool useAgentic

169.6

Elo · rank #371

Parameters

2.7B

Active params

2.7B (dense)

Context

4K tokens

Architecture

Decoder-only Transformer, deep-and-thin design (40 layers, 2304 hidden dim, Mistral-derived tokenizer)

License

Custom (OpenBMB General Model License; free for research, free commercial use after registration)

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning68.0%#57

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning85.4%#13

best: Phi-3-medium (14B) · 97.7%

BIG-Bench HardReasoning36.9%#117

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge51.1%#50

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge51.1%#28

best: Doubao-1.5-Pro · 90.9%

GSM8KMath53.8%#127

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning68.3%#132

best: Claude 3 Opus · 95.4%

HumanEvalCoding50.0%#124

best: Claude Opus 4.5 · 99.4%

MATH-500Math10.2%#193

best: GPT-5 · 99.4%

MBPPCoding47.3%#76

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLUKnowledge53.5%#227

best: OpenAI o3 · 92.9%

MT-BenchHuman preference7.25#56

best: Hunyuan-Large (A52B) · 9.4

Run it locally

VRAM @ Q4

2 GB

VRAM @ FP16

6.2 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF · Int4

MiniCPM family

Elo progression across releases

MiniCPM-2BFeb 2024169.6

Resources

HFopenbmb/MiniCPM-2B-sft-bf16 GHOpenBMB/MiniCPM

API price weights · each benchmark row carries its own source badge (see methodology)