Improve model card: Add paper/project links and `code-generation`, `datasets`, `metrics` tags
Browse filesThis PR enhances the model card for MiniCPM4-8B by:
* Adding direct links to the official Hugging Face paper page ([MiniCPM4: Ultra-Efficient LLMs on End Devices](https://huggingface.co/papers/2506.07900)) and the Hugging Face collection project page (`https://huggingface.co/collections/openbmb/minicpm4-6841ab29d180257e940baa9b`) in the header section. Existing links pointing to the GitHub technical report PDF in "What's New" and "Citation" are updated to the HF paper link. The BibTeX entry is also updated with the arXiv journal.
* Introducing the `code-generation` tag to the metadata, as supported by the model's performance on code benchmarks (HumanEval+, MBPP+, LiveCodeBench) and the "MCP-Code-Executor" application mentioned in the GitHub README.
* Adding the `datasets` tag for `openbmb/Ultra-FineWeb` and the `metrics` tag for `accuracy`, as both are explicitly mentioned and demonstrated in the model's documentation.
* Restructuring the "Usage" section to align with the original GitHub README's presentation of `InfLLM v2` and `rope_scaling` details under the Hugging Face inference, rather than the `CPM.cu` section. This ensures consistency and clarity for users.
These improvements will make the model more discoverable, accurately categorized, and provide a clearer usage guide on the Hugging Face Hub.
README.md
CHANGED
|
@@ -1,123 +1,348 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
<
|
| 17 |
-
<
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
- [
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
- [MiniCPM4-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
-
|
| 51 |
-
|
| 52 |
-
-
|
| 53 |
-
-
|
| 54 |
-
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
#
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- zh
|
| 4 |
+
- en
|
| 5 |
+
library_name: transformers
|
| 6 |
+
license: apache-2.0
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
tags:
|
| 9 |
+
- code-generation
|
| 10 |
+
datasets:
|
| 11 |
+
- openbmb/Ultra-FineWeb
|
| 12 |
+
metrics:
|
| 13 |
+
- accuracy
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
<div align="center">
|
| 17 |
+
<img src="https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm_logo.png?raw=true" width="500em" ></img>
|
| 18 |
+
</div>
|
| 19 |
+
|
| 20 |
+
<p align="center">
|
| 21 |
+
<a href="https://huggingface.co/papers/2506.07900" target="_blank">Paper</a> |
|
| 22 |
+
<a href="https://github.com/OpenBMB/MiniCPM/" target="_blank">GitHub Repo</a> |
|
| 23 |
+
<a href="https://huggingface.co/collections/openbmb/minicpm4-6841ab29d180257e940baa9b" target="_blank">Project Page</a> |
|
| 24 |
+
<a href="https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf" target="_blank">Technical Report</a>
|
| 25 |
+
</p>
|
| 26 |
+
<p align="center">
|
| 27 |
+
👋 Join us on <a href="https://discord.gg/3cGQn9b3YM" target="_blank">Discord</a> and <a href="https://github.com/OpenBMB/MiniCPM/blob/main/assets/wechat.jpg" target="_blank">WeChat</a>
|
| 28 |
+
</p>
|
| 29 |
+
|
| 30 |
+
## What's New
|
| 31 |
+
- [2025.06.06] **MiniCPM4** series are released! This model achieves ultimate efficiency improvements while maintaining optimal performance at the same scale! It can achieve over 5x generation acceleration on typical end-side chips! You can find the paper [here](https://huggingface.co/papers/2506.07900).🔥🔥🔥
|
| 32 |
+
|
| 33 |
+
## MiniCPM4 Series
|
| 34 |
+
MiniCPM4 series are highly efficient large language models (LLMs) designed explicitly for end-side devices, which achieves this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems.
|
| 35 |
+
- [MiniCPM4-8B](https://huggingface.co/openbmb/MiniCPM4-8B): The flagship of MiniCPM4, with 8B parameters, trained on 8T tokens. (**<-- you are here**)
|
| 36 |
+
- [MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B): The small version of MiniCPM4, with 0.5B parameters, trained on 1T tokens.
|
| 37 |
+
- [MiniCPM4-8B-Eagle-FRSpec](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec): Eagle head for FRSpec, accelerating speculative inference for MiniCPM4-8B.
|
| 38 |
+
- [MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu): Eagle head trained with QAT for FRSpec, efficiently integrate speculation and quantization to achieve ultra acceleration for MiniCPM4-8B.
|
| 39 |
+
- [MiniCPM4-8B-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-vLLM): Eagle head in vLLM format, accelerating speculative inference for MiniCPM4-8B.
|
| 40 |
+
- [MiniCPM4-8B-marlin-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-marlin-Eagle-vLLM): Quantized Eagle head for vLLM format, accelerating speculative inference for MiniCPM4-8B.
|
| 41 |
+
- [BitCPM4-0.5B](https://huggingface.co/openbmb/BitCPM4-0.5B): Extreme ternary quantization applied to MiniCPM4-0.5B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
|
| 42 |
+
- [BitCPM4-1B](https://huggingface.co/openbmb/BitCPM4-1B): Extreme ternary quantization applied to MiniCPM3-1B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
|
| 43 |
+
- [MiniCPM4-Survey](https://huggingface.co/openbmb/MiniCPM4-Survey): Based on MiniCPM4-8B, accepts users' quiries as input and autonomously generate trustworthy, long-form survey papers.
|
| 44 |
+
- [MiniCPM4-MCP](https://huggingface.co/openbmb/MiniCPM4-MCP): Based on MiniCPM4-8B, accepts users' queries and available MCP tools as input and autonomously calls relevant MCP tools to satisfy users' requirements.
|
| 45 |
+
|
| 46 |
+
## Introduction
|
| 47 |
+
MiniCPM 4 is an extremely efficient edge-side large model that has undergone efficient optimization across four dimensions: model architecture, learning algorithms, training data, and inference systems, achieving ultimate efficiency improvements.
|
| 48 |
+
|
| 49 |
+
- 🏗️ **Efficient Model Architecture:**
|
| 50 |
+
- InfLLM v2 -- Trainable Sparse Attention Mechanism: Adopts a trainable sparse attention mechanism architecture where each token only needs to compute relevance with less than 5% of tokens in 128K long text processing, significantly reducing computational overhead for long texts
|
| 51 |
+
|
| 52 |
+
- 🧠 **Efficient Learning Algorithms:**
|
| 53 |
+
- Model Wind Tunnel 2.0 -- Efficient Predictable Scaling: Introduces scaling prediction methods for performance of downstream tasks, enabling more precise model training configuration search
|
| 54 |
+
- BitCPM -- Ultimate Ternary Quantization: Compresses model parameter bit-width to 3 values, achieving 90% extreme model bit-width reduction
|
| 55 |
+
- Efficient Training Engineering Optimization: Adopts FP8 low-precision computing technology combined with Multi-token Prediction training strategy
|
| 56 |
+
|
| 57 |
+
- 📚 **High-Quality Training Data:**
|
| 58 |
+
- UltraClean -- High-quality Pre-training Data Filtering and Generation: Builds iterative data cleaning strategies based on efficient data verification, open-sourcing high-quality Chinese and English pre-training dataset [UltraFinweb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb)
|
| 59 |
+
- UltraChat v2 -- High-quality Supervised Fine-tuning Data Generation: Constructs large-scale high-quality supervised fine-tuning datasets covering multiple dimensions including knowledge-intensive data, reasoning-intensive data, instruction-following data, long text understanding data, and tool calling data
|
| 60 |
+
|
| 61 |
+
- ⚡ **Efficient Inference System:**
|
| 62 |
+
- CPM.cu -- Lightweight and Efficient CUDA Inference Framework: Integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding
|
| 63 |
+
- ArkInfer -- Cross-platform Deployment System: Supports efficient deployment across multiple backend environments, providing flexible cross-platform adaptation capabilities
|
| 64 |
+
|
| 65 |
+
## Usage
|
| 66 |
+
|
| 67 |
+
### Inference with [CPM.cu](https://github.com/OpenBMB/cpm.cu)
|
| 68 |
+
|
| 69 |
+
We recommend using [CPM.cu](https://github.com/OpenBMB/cpm.cu) for the inference of MiniCPM4. CPM.cu is a CUDA inference framework developed by OpenBMB, which integrates efficient sparse, speculative sampling, and quantization techniques, fully leveraging the efficiency advantages of MiniCPM4.
|
| 70 |
+
|
| 71 |
+
You can install CPM.cu by running the following command:
|
| 72 |
+
|
| 73 |
+
```bash
|
| 74 |
+
git clone https://github.com/OpenBMB/cpm.cu.git --recursive
|
| 75 |
+
cd cpm.cu
|
| 76 |
+
python3 setup.py install
|
| 77 |
+
```
|
| 78 |
+
|
| 79 |
+
After modification, you can run the following command to reproduce the long-context acceleration effect (the script will automatically download the model weights from HuggingFace)
|
| 80 |
+
```bash
|
| 81 |
+
python3 tests/long_prompt_gen.py # 生成 prompt.txt
|
| 82 |
+
python3 tests/test_generate.py --prompt-file prompt.txt
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
For more details about CPM.cu, please refer to [the repo CPM.cu](https://github.com/OpenBMB/cpm.cu).
|
| 86 |
+
|
| 87 |
+
### HuggingFace
|
| 88 |
+
|
| 89 |
+
```python
|
| 90 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 91 |
+
import torch
|
| 92 |
+
torch.manual_seed(0)
|
| 93 |
+
|
| 94 |
+
path = 'openbmb/MiniCPM4-8B'
|
| 95 |
+
device = "cuda"
|
| 96 |
+
tokenizer = AutoTokenizer.from_pretrained(path)
|
| 97 |
+
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
|
| 98 |
+
|
| 99 |
+
# User can directly use the chat interface
|
| 100 |
+
# responds, history = model.chat(tokenizer, "Write an article about Artificial Intelligence.", temperature=0.7, top_p=0.7)
|
| 101 |
+
# print(responds)
|
| 102 |
+
|
| 103 |
+
# User can also use the generate interface
|
| 104 |
+
messages = [
|
| 105 |
+
{"role": "user", "content": "Write an article about Artificial Intelligence."},
|
| 106 |
+
]
|
| 107 |
+
prompt_text = tokenizer.apply_chat_template(
|
| 108 |
+
messages,
|
| 109 |
+
tokenize=False,
|
| 110 |
+
add_generation_prompt=True,
|
| 111 |
+
)
|
| 112 |
+
model_inputs = tokenizer([prompt_text], return_tensors="pt").to(device)
|
| 113 |
+
|
| 114 |
+
model_outputs = model.generate(
|
| 115 |
+
**model_inputs,
|
| 116 |
+
max_new_tokens=1024,
|
| 117 |
+
top_p=0.7,
|
| 118 |
+
temperature=0.7
|
| 119 |
+
)
|
| 120 |
+
output_token_ids = [
|
| 121 |
+
model_outputs[i][len(model_inputs[i]):] for i in range(len(model_inputs['input_ids']))
|
| 122 |
+
]
|
| 123 |
+
|
| 124 |
+
responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]
|
| 125 |
+
print(responses)
|
| 126 |
+
```
|
| 127 |
+
|
| 128 |
+
本模型支持稀疏注意力机制 InfLLM v2，可高效处理长序列推理。如需启用该功能，请先安装依赖库 [infllmv2_cuda_impl](https://github.com/OpenBMB/infllmv2_cuda_impl)
|
| 129 |
+
|
| 130 |
+
运行以下命令即可安装：
|
| 131 |
+
|
| 132 |
+
```bash
|
| 133 |
+
git clone -b feature_infer https://github.com/OpenBMB/infllmv2_cuda_impl.git
|
| 134 |
+
cd infllmv2_cuda_impl
|
| 135 |
+
git submodule update --init --recursive
|
| 136 |
+
pip install -e . # or python setup.py install
|
| 137 |
+
```
|
| 138 |
+
|
| 139 |
+
启用 InfLLM v2 需在 `config.json` 配置文件中添加 `sparse_config` 字段：
|
| 140 |
+
|
| 141 |
+
```json
|
| 142 |
+
{
|
| 143 |
+
...,
|
| 144 |
+
"sparse_config": {
|
| 145 |
+
"kernel_size": 32,
|
| 146 |
+
"kernel_stride": 16,
|
| 147 |
+
"init_blocks": 1,
|
| 148 |
+
"block_size": 64,
|
| 149 |
+
"window_size": 2048,
|
| 150 |
+
"topk": 64,
|
| 151 |
+
"use_nope": false,
|
| 152 |
+
"dense_len": 8192
|
| 153 |
+
}
|
| 154 |
+
}
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
These parameters control InfLLM v2 的行为:
|
| 158 |
+
|
| 159 |
+
* `kernel_size`（默认值：32）：语义核的大小。
|
| 160 |
+
* `kernel_stride`（默认值：16）：相邻语义核的步长。
|
| 161 |
+
* `init_blocks`（默认值：1）：每个 query token 关注的初始的块数量，用于确保关注序列开头部分。
|
| 162 |
+
* `block_size`（默认值：64）：key-value blocks 的块大小。
|
| 163 |
+
* `window_size`（默认值：2048）：局部滑动窗口大小。
|
| 164 |
+
* `topk`（默认值：64）：每个 token 仅与最相关的 top-k 个 key-value blocks 计算注意力。
|
| 165 |
+
* `use_nope`（默认值：false）：是否在块选择中使用NOPE技术以提升性能。
|
| 166 |
+
* `dense_len`（默认值：8192）：稀疏注意力对短序列收益有限，当 token 长度低于此阈值时自动切换为标准注意力。设为 `-1` 则强制始终使用稀疏注意力。
|
| 167 |
+
|
| 168 |
+
Minicpm4 原生支持 32,768 tokens 的上下文长度。若对话总长度（输入 + 输出）远超此限制，建议通过 RoPE 缩放技术扩展上下文。我们已验证通过调整 LongRoPE 因子，模型可稳定支持 131,072 tokens 的超长上下文。
|
| 169 |
+
|
| 170 |
+
修改方法：在 `config.json` 文件中调整 `rope_scaling` 字段参数即可。
|
| 171 |
+
|
| 172 |
+
```json
|
| 173 |
+
{
|
| 174 |
+
...,
|
| 175 |
+
"rope_scaling": {
|
| 176 |
+
"rope_type": "longrope",
|
| 177 |
+
"long_factor": [
|
| 178 |
+
0.9977997200264581,
|
| 179 |
+
1.014658295992452,
|
| 180 |
+
1.0349680404997148,
|
| 181 |
+
1.059429246056193,
|
| 182 |
+
1.0888815016813513,
|
| 183 |
+
1.1243301355211495,
|
| 184 |
+
1.166977103606075,
|
| 185 |
+
1.2182568066927284,
|
| 186 |
+
1.2798772354275727,
|
| 187 |
+
1.3538666751582975,
|
| 188 |
+
1.4426259039919596,
|
| 189 |
+
1.5489853358570191,
|
| 190 |
+
1.6762658237220625,
|
| 191 |
+
1.8283407612492941,
|
| 192 |
+
2.0096956085876183,
|
| 193 |
+
2.225478927469756,
|
| 194 |
+
2.481536379650452,
|
| 195 |
+
2.784415934557119,
|
| 196 |
+
3.1413289096347365,
|
| 197 |
+
3.560047844772632,
|
| 198 |
+
4.048719380066383,
|
| 199 |
+
4.615569542115128,
|
| 200 |
+
5.2684819496549835,
|
| 201 |
+
6.014438591970396,
|
| 202 |
+
6.858830049237097,
|
| 203 |
+
7.804668263503327,
|
| 204 |
+
8.851768731513417,
|
| 205 |
+
9.99600492938444,
|
| 206 |
+
11.228766118181639,
|
| 207 |
+
12.536757560834843,
|
| 208 |
+
13.902257701387796,
|
| 209 |
+
15.303885189125953,
|
| 210 |
+
16.717837610115794,
|
| 211 |
+
18.119465097853947,
|
| 212 |
+
19.484965238406907,
|
| 213 |
+
20.792956681060105,
|
| 214 |
+
22.02571786985731,
|
| 215 |
+
23.16995406772833,
|
| 216 |
+
24.217054535738416,
|
| 217 |
+
25.16289275000465,
|
| 218 |
+
26.007284207271347,
|
| 219 |
+
26.753240849586767,
|
| 220 |
+
27.40615325712662,
|
| 221 |
+
27.973003419175363,
|
| 222 |
+
28.461674954469114,
|
| 223 |
+
28.880393889607006,
|
| 224 |
+
29.237306864684626,
|
| 225 |
+
29.540186419591297,
|
| 226 |
+
29.79624387177199,
|
| 227 |
+
30.01202719065413,
|
| 228 |
+
30.193382037992453,
|
| 229 |
+
30.34545697551969,
|
| 230 |
+
30.47273746338473,
|
| 231 |
+
30.579096895249787,
|
| 232 |
+
30.66785612408345,
|
| 233 |
+
30.741845563814174,
|
| 234 |
+
30.80346599254902,
|
| 235 |
+
30.85474569563567,
|
| 236 |
+
30.897392663720595,
|
| 237 |
+
30.932841297560394,
|
| 238 |
+
30.962293553185553,
|
| 239 |
+
30.986754758742034,
|
| 240 |
+
31.007064503249293,
|
| 241 |
+
31.02392307921529
|
| 242 |
+
],
|
| 243 |
+
"short_factor": [
|
| 244 |
+
0.9977997200264581,
|
| 245 |
+
1.014658295992452,
|
| 246 |
+
1.0349680404997148,
|
| 247 |
+
1.059429246056193,
|
| 248 |
+
1.0888815016813513,
|
| 249 |
+
1.1243301355211495,
|
| 250 |
+
1.166977103606075,
|
| 251 |
+
1.2182568066927284,
|
| 252 |
+
1.2798772354275727,
|
| 253 |
+
1.3538666751582975,
|
| 254 |
+
1.4426259039919596,
|
| 255 |
+
1.5489853358570191,
|
| 256 |
+
1.6762658237220625,
|
| 257 |
+
1.8283407612492941,
|
| 258 |
+
2.0096956085876183,
|
| 259 |
+
2.225478927469756,
|
| 260 |
+
2.481536379650452,
|
| 261 |
+
2.784415934557119,
|
| 262 |
+
3.1413289096347365,
|
| 263 |
+
3.560047844772632,
|
| 264 |
+
4.048719380066383,
|
| 265 |
+
4.615569542115128,
|
| 266 |
+
5.2684819496549835,
|
| 267 |
+
6.014438591970396,
|
| 268 |
+
6.858830049237097,
|
| 269 |
+
7.804668263503327,
|
| 270 |
+
8.851768731513417,
|
| 271 |
+
9.99600492938444,
|
| 272 |
+
11.228766118181639,
|
| 273 |
+
12.536757560834843,
|
| 274 |
+
13.902257701387796,
|
| 275 |
+
15.303885189125953,
|
| 276 |
+
16.717837610115794,
|
| 277 |
+
18.119465097853947,
|
| 278 |
+
19.484965238406907,
|
| 279 |
+
20.792956681060105,
|
| 280 |
+
22.02571786985731,
|
| 281 |
+
23.16995406772833,
|
| 282 |
+
24.217054535738416,
|
| 283 |
+
25.16289275000465,
|
| 284 |
+
26.007284207271347,
|
| 285 |
+
26.753240849586767,
|
| 286 |
+
27.40615325712662,
|
| 287 |
+
27.973003419175363,
|
| 288 |
+
28.461674954469114,
|
| 289 |
+
28.880393889607006,
|
| 290 |
+
29.237306864684626,
|
| 291 |
+
29.540186419591297,
|
| 292 |
+
29.79624387177199,
|
| 293 |
+
30.01202719065413,
|
| 294 |
+
30.193382037992453,
|
| 295 |
+
30.34545697551969,
|
| 296 |
+
30.47273746338473,
|
| 297 |
+
30.579096895249787,
|
| 298 |
+
30.66785612408345,
|
| 299 |
+
30.741845563814174,
|
| 300 |
+
30.80346599254902,
|
| 301 |
+
30.85474569563567,
|
| 302 |
+
30.897392663720595,
|
| 303 |
+
30.932841297560394,
|
| 304 |
+
30.962293553185553,
|
| 305 |
+
30.986754758742034,
|
| 306 |
+
31.007064503249293,
|
| 307 |
+
31.02392307921529
|
| 308 |
+
]
|
| 309 |
+
},
|
| 310 |
+
"original_max_position_embeddings": 32768
|
| 311 |
+
}
|
| 312 |
+
```
|
| 313 |
+
|
| 314 |
+
## Evaluation Results
|
| 315 |
+
On two typical end-side chips, Jetson AGX Orin and RTX 4090, MiniCPM4 demonstrates significantly faster processing speed compared to similar-size models in long text processing tasks. As text length increases, MiniCPM4's efficiency advantage becomes more pronounced. On the Jetson AGX Orin platform, compared to Qwen3-8B, MiniCPM4 achieves approximately 7x decoding speed improvement.
|
| 316 |
+
|
| 317 |
+
|
| 318 |
+
|
| 319 |
+
#### Comprehensive Evaluation
|
| 320 |
+
MiniCPM4 launches end-side versions with 8B and 0.5B parameter scales, both achieving best-in-class performance in their respective categories.
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
|
| 324 |
+
#### Long Text Evaluation
|
| 325 |
+
MiniCPM4 is pre-trained on 32K long texts and achieves length extension through YaRN technology. In the 128K long text needle-in-a-haystack task, MiniCPM4 demonstrates outstanding performance.
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
## Statement
|
| 330 |
+
- As a language model, MiniCPM generates content by learning from a vast amount of text.
|
| 331 |
+
- However, it does not possess the ability to comprehend or express personal opinions or value judgments.
|
| 332 |
+
- Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers.
|
| 333 |
+
- Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.
|
| 334 |
+
|
| 335 |
+
## LICENSE
|
| 336 |
+
- This repository and MiniCPM models are released under the [Apache-2.0](https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE) License.
|
| 337 |
+
|
| 338 |
+
## Citation
|
| 339 |
+
- Please cite our [paper](https://huggingface.co/papers/2506.07900) if you find our work valuable.
|
| 340 |
+
|
| 341 |
+
```bibtex
|
| 342 |
+
@article{minicpm4,
|
| 343 |
+
title={{MiniCPM4}: Ultra-Efficient LLMs on End Devices},
|
| 344 |
+
author={MiniCPM Team},
|
| 345 |
+
year={2025},
|
| 346 |
+
journal={arXiv preprint arXiv:2506.07900}
|
| 347 |
+
}
|
| 348 |
+
```
|