Model

MiniCPM3-4B

OPEN

OpenBMB · MiniCPM 3 family · released Sep 5, 2024

Third-gen 4B on-device LLM with function calling and code interpreter, outperforming GPT-3.5-Turbo-0125 and Phi-3.5-mini.

ReasoningCodingVisionFunction callingTool useAgentic

579.8

Elo · rank #284

Parameters

Active params

4B (dense)

Context

32K tokens

Architecture

Dense causal transformer (third-gen MiniCPM)

License

MiniCPM Model License (custom; free for research and, after registration, commercial use) + Apache-2.0 code

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

BIG-Bench HardReasoning70.2%#57

best: ERNIE 4.5 300B-A47B · 94.3%

BFCLAgents76.0%#17

best: SmolLM3 3B (No Thinking) · 92.3%

C-EvalKnowledge73.6%#38

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge73.3%#18

best: Doubao-1.5-Pro · 90.9%

GSM8KMath81.1%#77

best: Llama 3.1 405B · 96.8%

HumanEval+Coding68.3%#19

best: Mistral Small 3.2 (24B) · 92.9%

IFEvalReasoning68.4%#117

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding22.6%#162

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math46.6%#140

best: GPT-5 · 99.4%

MBPP+Coding63.2%#25

best: Llama 3.1 405B · 88.6%

MMLUKnowledge67.2%#168

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.41#29

best: Hunyuan-Large (A52B) · 9.4

Run it locally

VRAM @ Q4

3 GB

VRAM @ FP16

8 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 Q4_K_M · GGUF Q5_K_M · GGUF Q8_0

MiniCPM 3 family

Elo progression across releases

MiniCPM3-4BSep 2024579.8

Resources

HFopenbmb/MiniCPM3-4B GHOpenBMB/MiniCPM

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