MiniCPM Tech Report | MiniCPM Wiki (in Chinese) | MiniCPM-V Repo | UltraData
Join our discord and Feishu/Lark | Join Us
Note
The MiniCPM-SALA architecture is just the beginning. Realizing its full potential requires deep system-level synergy and cross-layer compilation optimization.
OpenBMB, in collaboration with SGLang and NVIDIA, invites global geeks to tackle the limits of 9B-scale, 1M-token inference on a dedicated NVIDIA 6000D environment.
- 💰 Prize Pool: >$100,000 USD (Top Prize: $89,000)
- 🚀 Goal: Optimize single and multi-batch performance via cross-layer compilation.
We are releasing MiniCPM5-1B, the first model in the MiniCPM5 series. It is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios, reaching 1B-class open-source SOTA.
🏆 1B-class open-source SOTA: MiniCPM5-1B reaches an average score of 42.57 across reasoning, knowledge, code, instruction-following, math, logic and agentic benchmarks, above the highest average score of 35.61 among strong open-source models in the same size class; its strengths are most visible in agentic tool use, code, and competition math.
🧠 Hybrid Reasoning: built-in <think> chat template, switch via enable_thinking. The same checkpoint serves as both a fast assistant and a deliberate reasoner.
🛠️ Deployment / Fine-tuning Agent Skills: the repo provides single-page cookbooks for major inference backends and fine-tuning frameworks, each paired with an Agent Skill to help developers reproduce deployment and fine-tuning workflows.
🐱 Desktop Pet: a local-LLM desktop pet driven by MiniCPM5-1B — see Desktop Pet below.
- 📌 [2026.05.19] MiniCPM5-1B is released: a compact 1B-class dense model for on-device and resource-constrained use, paired with deployment / fine-tuning Agent Skills.
- [2026.02.11] MiniCPM-SALA is released: a sparse-and-linear hybrid attention model for million-token context modeling and efficient inference.
- [2025.09.05] MiniCPM4.1 series is released: a trainable sparse-attention model with hybrid reasoning.
- [2025.06.06] MiniCPM4 is released: an end-side model with over 5x generation acceleration on typical edge chips.
Older entries (2024 + InfLLM-V2 paper)
- [2025.09.29] InfLLM-V2 paper is released! We can train a sparse attention model with only 5B long-text tokens.
- [2024.09.05] We release MiniCPM3-4B! This model outperforms Phi-3.5-mini-instruct and GPT-3.5-Turbo-0125 and is comparable to several models with 7B-9B parameters like Llama3.1-8B-Instruct, Qwen2-7B-Instruct, and GLM-4-9B-Chat.
- [2024.07.05] Released MiniCPM-S-1B! This model achieves an average sparsity of 87.89% in the FFN layer, reducing FFN FLOPs by 84%, while maintaining downstream task performance.
- [2024.04.11] Released MiniCPM-2B-128k, MiniCPM-MoE-8x2B and MiniCPM-1B! Click here to read our technical blog.
- [2024.02.01] Released MiniCPM-2B! This model performs similarly to Mistral-7B on public benchmarks (with better performance in Chinese, math, and code abilities) and overall outperforms models like Llama2-13B, MPT-30B, and Falcon-40B.
- ✨ Highlights
- 🔥 Changelog
- 📦 Model Downloads
- 🚀 MiniCPM5-1B
- 🧪 MiniCPM-SALA
- ⚡ MiniCPM4 & MiniCPM4.1 Series
- Legacy topics →: BitCPM4 quantization, MiniCPM4 applications
- 📄 LICENSE · 🏛 Institutions · 📚 Citation
Current release: MiniCPM5-1B (BF16, GGUF, MLX):
| HuggingFace | ModelScope |
|---|---|
| MiniCPM5-1B | MiniCPM5-1B |
| MiniCPM5-1B-SFT | MiniCPM5-1B-SFT |
| MiniCPM5-1B-Base | MiniCPM5-1B-Base |
| MiniCPM5-1B-GGUF | MiniCPM5-1B-GGUF |
| MiniCPM5-1B-MLX | MiniCPM5-1B-MLX |
Other key releases:
| HuggingFace | ModelScope |
|---|---|
| MiniCPM-SALA | MiniCPM-SALA |
| MiniCPM4.1-8B | MiniCPM4.1-8B |
| MiniCPM4-0.5B | MiniCPM4-0.5B |
📋 Click to view earlier MiniCPM releases: 4, BitCPM, applications, MiniCPM3 / 2B / 1B
Earlier flagships:
| HuggingFace | ModelScope |
|---|---|
| MiniCPM4-8B | MiniCPM4-8B |
MiniCPM4.1 quantized & speculative variants:
BitCPM4 ternary-quantized + MiniCPM4 Applications:
| HuggingFace | ModelScope |
|---|---|
| BitCPM4-1B | BitCPM4-1B |
| BitCPM4-0.5B | BitCPM4-0.5B |
| MiniCPM4-Survey | MiniCPM4-Survey |
| MiniCPM4-MCP | MiniCPM4-MCP |
MiniCPM4 Eagle speculative decoding, QAT, and pre-2025 releases:
MiniCPM5-1B is the first checkpoint in the MiniCPM5 series. It is designed for local assistants, coding agents, tool-use workflows, and reasoning scenarios where a compact model is preferred. The model keeps a small deployment footprint while providing native long-context support and both Think / No Think chat modes through the same checkpoint.
We compare MiniCPM5-1B with strong open-source models in the same size class, including LFM2.5-1.2B-Thinking, Qwen3-0.6B/think and Qwen3.5-0.8B/think. These are capable baselines; within this comparison set, MiniCPM5-1B reaches 1B-class open-source SOTA, with its advantage most visible in tool use, code generation, and difficult reasoning. This makes it a practical choice for local coding agents, tool assistants, and reasoning assistants.
The training of MiniCPM5-1B is a full-stack practice of UltraData Tiered Data Management, covering three stages: base training, mid-training, and post-training.
During base training, the model goes through stable training and decay training to build core language capability and training stability. It then enters mid-training to further strengthen target capabilities and adapt to the target data distribution. The training corpus is released alongside the model as Ultra-FineWeb, Ultra-FineWeb-L3, and UltraData-Math.
During post-training, we proceed in three steps: SFT, RL, and OPD. We first use 200B tokens of deep-thinking SFT and 200B tokens of hybrid-thinking SFT to establish deep-thinking, hybrid-thinking, and general chat abilities; the SFT data is released as UltraData-SFT-2605. We then train specialized RL teachers for math, code, closed-book QA, writing, and related domains, and use On-Policy Distillation (OPD) to distill these teachers back into one release model.
RL + OPD is a key part of MiniCPM5-1B post-training. On math, code and instruction-following tasks, RL + OPD raises the average score by ↑16 points while cutting the share of responses that hit the max-tokens budget by ↓29 percentage points. The figures below show the two-stage Reasoning RL pipeline, score gains, and the drop in overlong responses.
RL combines complementary training signals for reasoning, closed-book QA, writing, instruction following, long-context understanding, and general dialogue. Reasoning RL is based on DAPO-Math-17k (inspired by JustRL's minimalist recipe) and uses a two-stage length schedule to reduce overlong responses while improving reasoning accuracy. We also use TriviaQA, NQ-Open, LongWriter-Zero-RLData, synthesized verifiable RLVR data, and pair-wise RLHF signals to improve reliability, instruction following, and user experience.
OPD builds on Thinking Machines Lab's On-Policy Distillation and incorporates implementation improvements from Rethinking On-Policy Distillation. In the RL framework, we use reverse KL divergence as the advantage estimate, replacing the original verification-based advantage. At each response position, we take top-k logits from both the student and teacher models, compute reverse KL on the union of the two token sets, and balance the accuracy of the RKL signal with training efficiency. OPD reuses the in-domain prompts used to train each RL teacher as distillation data, so no additional data curation is required.
For the three most common inference paths, you can start a service or run local inference as follows:
vLLM (OpenAI-compatible server, NVIDIA GPU):
pip install "vllm>=0.21" && vllm serve openbmb/MiniCPM5-1B --port 8000Test request:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "openbmb/MiniCPM5-1B",
"messages": [{"role": "user", "content": "Who are you? Please briefly introduce yourself."}],
"max_tokens": 128,
"temperature": 0.7
}'SGLang (OpenAI-compatible server, NVIDIA GPU):
pip install "sglang[srt]>=0.5.12" && python -m sglang.launch_server --model-path openbmb/MiniCPM5-1B --port 30000Test request:
curl http://localhost:30000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "openbmb/MiniCPM5-1B",
"messages": [{"role": "user", "content": "Who are you? Please briefly introduce yourself."}],
"max_tokens": 128,
"temperature": 0.7
}'Transformers (local Python inference, CPU or GPU):
pip install -U "transformers>=5.6" accelerate torchLocal inference test:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "openbmb/MiniCPM5-1B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
messages = [{"role": "user", "content": "Who are you? Please briefly introduce yourself."}]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
enable_thinking=False,
return_dict=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))Recommended chat template sampling:
| Mode | Recommended params | Enable |
|---|---|---|
| Think | temperature=0.9, top_p=0.95 |
enable_thinking=True |
| No Think | temperature=0.7, top_p=0.95 |
enable_thinking=False |
For other backends (llama.cpp / Ollama / LM Studio / MLX), see the cookbooks table below.
For tool / function calling, SGLang is the recommended backend — MiniCPM5-1B emits XML-style tool calls and SGLang's built-in minicpm5 parser converts them to OpenAI-compatible tool_calls natively:
python -m sglang.launch_server --model-path openbmb/MiniCPM5-1B --port 30000 \
--tool-call-parser minicpm5 # or: --tool-call-parser autoMiniCPM5-1B uses the standard LlamaForCausalLM architecture, so mainstream inference engines can load it directly: no custom kernels, no model-code fork. To help developers reproduce deployment and fine-tuning workflows, we provide single-page cookbooks paired with Cursor Agent Skills: the cookbooks are for manual execution, while Agent Skills let Cursor / Claude Code agents choose the route based on the target backend, hardware, and data path.
The two top-level skills cover deployment and fine-tuning:
| Top-level skill | What it does | Routes to |
|---|---|---|
minicpm5-deploy |
Inference router | transformers · vllm · sglang · llama-cpp · ollama · lmstudio · mlx |
minicpm5-finetune |
Fine tuning router | trl · llamafactory · ms-swift · unsloth · xtuner |
In Cursor / Claude Code, you can call them like this: the agent reads the top-level skill, selects the matching sub-skill and cookbook based on the target backend, hardware, and data path, then runs the command and reports back.
@minicpm5-deploy serve openbmb/MiniCPM5-1B with vLLM on port 8000
@minicpm5-finetune use unsloth + LoRA on /data/my_chat.jsonl, write to ./out
The tables below list the cookbook and sub-skill for each inference backend and fine-tuning framework. Quantized models are not listed as standalone backends; they are described under the inference backend that can load each format.
Inference Deployment (7 backends)
| Backend | Model format / use case | Cookbook | Paired Agent Skill |
|---|---|---|---|
| Transformers | BF16 / FP16 local Python inference, GPU + CPU | docs/deployment/transformers.md |
minicpm5-deploy-transformers |
| SGLang | BF16 / FP16 OpenAI server, recommended for tool calling | docs/deployment/sglang.md |
minicpm5-deploy-sglang |
| llama.cpp | GGUF local inference, CPU/GPU | docs/deployment/llama_cpp.md |
minicpm5-deploy-llama-cpp |
| Ollama | GGUF local on-device runtime | docs/deployment/ollama.md |
minicpm5-deploy-ollama |
| LM Studio | GGUF Mac desktop app and OpenAI server | docs/deployment/lmstudio.md |
minicpm5-deploy-lmstudio |
| MLX | MLX / 4bit local inference on Apple Silicon | docs/deployment/mlx.md |
minicpm5-deploy-mlx |
| ArcLight | GGUF local on-device, CPU, Desktop & Server | docs/deployment/arclight.md |
minicpm5-deploy-arclight |
Fine tuning (5 frameworks)
In addition to the deployment and fine-tuning frameworks listed above, MiniCPM5-1B is also supported by FlagOS for multi-chip deployment.
To enable large-scale deployment across different AI chips, Beijing Zhiyuan Research Institute, together with numerous research institutions, chip manufacturers, system vendors, and algorithm and software organizations both domestically and internationally, jointly initiated and established the FlagOS Open Source Community.
The FlagOS community is dedicated to building a unified, open-source system software stack for various AI chips, encompassing core open-source projects such as a large-scale operator library, a unified AI compiler, parallel training and inference frameworks, and a unified communication library. It aims to create an open technology ecosystem connecting the “model-system-chip” layers. By enabling “develop once, deploy across chips”, FlagOS unlocks the computational potential of hardware, breaks down the ecosystem silos between different chip software stacks, and effectively reduces migration costs for developers.The FlagOS community fosters an AI hardware and software ecosystem, overcomes single-vendor closed-source monopolies, promotes widespread deployment of AI hardware technologies, and is committed to rooted in China while embracing global collaboration.
Official website express: https://flagos.io
FlagOS multi-chip support and usage
Thanks to FlagOS’s unified multi-chip AI system software stack, MiniCPM5-1B was adapted to 4–5 different AI chips in an extremely short time. Currently, the multi-chip version of MiniCPM5-1B has been released on FlagRelease, FlagOS’s platform for automatic migration, adaptation, and deployment of large models across multi-architecture AI chips. Details are as follows:
| Vendor | ModelScope | Huggingface |
|---|---|---|
| Nvidia | MiniCPM5-1B-nvidia-FlagOS | MiniCPM5-1B-nvidia-FlagOS |
| Hygon | MiniCPM5-1B-hygon-FlagOS | MiniCPM5-1B-hygon-FlagOS |
| Metax | MiniCPM5-1B-metax-FlagOS | MiniCPM5-1B-metax-FlagOS |
| Iluvatar | MiniCPM5-1B-iluvatar-FlagOS | MiniCPM5-1B-iluvatar-FlagOS |
| Zhenwu | MiniCPM5-1B-zhenwu-FlagOS | MiniCPM5-1B-zhenwu-FlagOS |
| Mthreads | MiniCPM5-1B-mthreads-FlagOS | MiniCPM5-1B-mthreads-FlagOS |
| Kunlunxin | MiniCPM5-1B-kunlunxin-FlagOS | MiniCPM5-1B-kunlunxin-FlagOS |
| Ascend | MiniCPM5-1B-ascend-FlagOS | MiniCPM5-1B-ascend-FlagOS |
| ARM-v9 | MiniCPM5-1B-Armv9-FlagOS | MiniCPM5-1B-Armv9-FlagOS |
FlagRelease is a platform developed by the FlagOS team for automatic migration, adaptation, and deployment of large models across multi-architecture AI chips. The multi-chip version of MiniCPM5-1B has already been released on FlagRelease. All necessary software packages are pre-installed on the platform, so users do not need to install anything.
| Vendor | ModelScope | Huggingface |
|---|---|---|
| Nvidia | MiniCPM5-1B-nvidia-FlagOS | MiniCPM5-1B-nvidia-FlagOS |
| Hygon | MiniCPM5-1B-hygon-FlagOS | MiniCPM5-1B-hygon-FlagOS |
| Metax | MiniCPM5-1B-metax-FlagOS | MiniCPM5-1B-metax-FlagOS |
| Iluvatar | MiniCPM5-1B-iluvatar-FlagOS | MiniCPM5-1B-iluvatar-FlagOS |
| Zhenwu | MiniCPM5-1B-zhenwu-FlagOS | MiniCPM5-1B-zhenwu-FlagOS |
| Mthreads | MiniCPM5-1B-mthreads-FlagOS | MiniCPM5-1B-mthreads-FlagOS |
| Kunlunxin | MiniCPM5-1B-kunlunxin-FlagOS | MiniCPM5-1B-kunlunxin-FlagOS |
| Ascend | MiniCPM5-1B-ascend-FlagOS | MiniCPM5-1B-ascend-FlagOS |
| ARM-v9 | MiniCPM5-1B-Armv9-FlagOS | MiniCPM5-1B-Armv9-FlagOS |
- Dependencies: Python 3.12, GLIBC 2.39, GLIBCXX 3.4.33, CXXABI 1.3.15
Official Repository: https://github.com/flagos-ai/FlagGems
pip install flag-gems==4.2.1rc0
pip install triton==3.5.1You can enable flagGems acceleration by adding the import of flagGems in the source code of vllm where inference is performed.
import flag_gems
flag_gems.enable(record=True, once=True, path="/root/gems.txt")vllm serve ${model_path} \
--trust-remote-code \
--dtype bfloat16 \
--enforce-eager \
--port ${Port} \
--served-model-name ${model_name} \
--gpu-memory-utilization 0.85vllm-plugin-FL is a plugin built for the vLLM inference/service framework. Developed on top of FlagOS’s unified multi-chip backend, it is designed to extend vLLM’s capabilities and performance across a variety of hardware environments.
| Vendor | From Scratch | From FlagRelease | |
|---|---|---|---|
| Nvidia | vllm-plugin-FL/MiniCPM5-1B | MiniCPM5-1B-ModelScope | MiniCPM5-1B-nvidia-FlagOS |
We also ship OpenBMB/MiniCPM-Desk-Pet, a desktop pet driven locally by MiniCPM5-1B. It uses a thin llama.cpp llama-server sidecar to load the GGUF model and serves an OpenAI-compatible local endpoint to an Electron pet UI.
The pet supports Apple Silicon / NVIDIA GPU / CPU paths, can work with coding agents such as Cursor, Claude Code, and Codex, and supports LoRA persona switching.
- User install: grab
Clawd-on-Desk-*-arm64.dmgfrom Releases, then follow the onboarding flow for environment checks, model download, and sidecar startup. - Developer run:
git clone git@github.com:OpenBMB/MiniCPM-Desk-Pet.git && ./go.sh— seeMiniCPM-Desk-Pet/README.mdfor the full setup.
The pet UI layer is forked from @rullerzhou-afk/clawd-on-desk (AGPL-3.0). On top of the upstream pet runtime, animation packs, and multi-agent integrations, we add the local MiniCPM5-1B sidecar, onboarding flow, and LoRA persona switching. Full attribution in
NOTICE.md.
First large-scale Sparse + Linear Attention hybrid for million-token context (2026-02). HuggingFace · ModelScope
Click to expand: highlights, evaluation, inference setup
MiniCPM-SALA (Sparse Attention and Linear Attention) is the first large-scale hybrid model effectively integrating sparse and linear attention for million-token context modeling
✅ Hybrid Architecture: combines 25% Sparse Attention (InfLLM-v2) for long-context modeling with 75% Linear Attention (Lightning Attention) for global efficiency.
✅ Inference Efficiency: achieves 3.5× inference speed and lower KV-cache overhead compared with dense baselines.
✅ Million-Token Context: uses HyPE (Hybrid Positional Embedding) to scale to 1M+ tokens while maintaining length generalization.
✅ HALO Adaptation: uses Hybrid Attention via Layer Optimization (HALO), a distillation recipe that transfers dense attention capabilities to the hybrid architecture and mitigates the degradation often seen in pure linear models.
MiniCPM-SALA is an efficient hybrid model in which 25% of the layers adopt InfLLM-V2 and the remaining 75% utilize Lightning Attention. This architecture enables inference of one million tokens on consumer GPUs such as the NVIDIA RTX 5090.
-
SALA Hybrid Attention Mechanism
- Integrates 25% InfLLM-V2 and 75% Lightning Attention, effectively leveraging the granular focus of sparse attention for local details and the high efficiency of linear attention for broad context.
-
Transformer-to-Hybrid Continue Training
- Circumvents the inefficiencies of cold-start training by performing an architectural transformation on the pre-trained weights, thereby reducing the total training budget to approximately 25% relative to training a comparable model from scratch.
-
HyPE (Hybrid Positional Encoding)
- Balances short-context and long-context performance, maintaining general capabilities (e.g., knowledge, mathematics, and coding) close to full-attention models such as Qwen3-8B while scoring higher on multiple long-context benchmarks.
-
Efficient Inference on Long Sequences
- Achieves up to 3.5x the inference speed of Qwen3-8B at a sequence length of 256K tokens on A6000D, supports inference at context lengths of up to 1M tokens on both NVIDIA A6000D and 5090 GPUs, whereas Qwen3-8B fails at this length due to out-of-memory (OOM) errors.
We benchmarked MiniCPM-SALA (9B) against Qwen3-8B on NVIDIA A6000D and RTX 5090 GPUs to evaluate inference speed and memory efficiency. MiniCPM-SALA achieves up to a 2.5x speedup in time-to-first-token (TTFT) and reduces the memory pressure of full-attention architectures at ultra-long lengths. In this setup, Qwen3-8B runs into OOM errors at extended lengths, while MiniCPM-SALA can process 1M-token contexts on a single consumer-grade RTX 5090.
MiniCPM-SALA scores higher than the tested open-source LLMs of similar scale on most long-context benchmarks. It achieves the highest scores in the RULER and NoLiMa tests at all context lengths up to 128K, with an overall average score of 38.97.
MiniCPM-SALA shows effective length extrapolation, maintaining a score of 81.6 at a 2048K context length despite being trained on up to 520K tokens. The model does this without auxiliary techniques like YaRN, likely due to its NoPE configuration in sparse attention layers.
MiniCPM-SALA achieves an average score of 76.53 across standard benchmarks, outperforming comparable models such as Qwen3-8B and Falcon-H1R-7B. The architecture maintains robust performance in Knowledge, Code, and Math.
Recommended inference setting: Temperature=0.9.
Our model is readily compatible with 🤗 Hugging Face transformers. You can perform inference with our model as follows:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_path = "openbmb/MiniCPM-SALA"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, device_map="auto")
model.eval()
prompts = ["My name is", "The capital of China is"]
with torch.no_grad():
inputs = tokenizer(prompts, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs)
output_texts = tokenizer.batch_decode(outputs)
print(output_texts)- CUDA 12.x or higher
gcc/g++compileruvpackage manager (script will check)
# Clone repository
git clone -b minicpm_sala https://github.com/OpenBMB/sglang.git
cd sglang
# One-click installation (creates venv and compiles all dependencies)
bash install_minicpm_sala.sh
# Or specify PyPI mirror
bash install_minicpm_sala.sh https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simpleThe installation script performs the following steps:
- Creates
sglang_minicpm_sala_envvirtual environment (Python 3.12) - Clones dependencies to
3rdparty/(infllmv2) and initializes submodules (sparse_kernel) - Installs MiniCPM-SALA (current repo)
- Compiles and installs
infllmv2_cuda_impl - Compiles and installs
sparse_kernel - Installs
tilelang&flash-linear-attention
# Activate environment
source sglang_minicpm_sala_env/bin/activate
# Launch Inference Server (Replace MODEL_PATH with actual path)
MODEL_PATH=/path/to/your/MiniCPM-SALA
python3 -m sglang.launch_server \
--model ${MODEL_PATH} \
--trust-remote-code \
--disable-radix-cache \
--attention-backend minicpm_flashinfer \
--chunked-prefill-size 8192 \
--max-running-requests 32 \
--skip-server-warmup \
--port 31111 \
--dense-as-sparse| Parameter | Description |
|---|---|
--trust-remote-code |
Allow custom code in model |
--disable-radix-cache |
Disable RadixAttention prefix cache |
--attention-backend minicpm_flashinfer |
Use MiniCPM FlashInfer backend |
--chunked-prefill-size 8192 |
Chunked prefill size |
--max-running-requests 32 |
Max concurrent requests |
--skip-server-warmup |
Skip server warmup |
--port 31111 |
Server port |
--dense-as-sparse |
Use dense-as-sparse mode |
If the script doesn't work for you, follow these steps:
# 0. Ensure uv is installed
pip install uv
# 1. Create venv
uv venv --python 3.12 sglang_minicpm_sala_env
source sglang_minicpm_sala_env/bin/activate
# 2. Install SGLang
uv pip install --upgrade pip setuptools wheel
uv pip install -e ./python[all]
# 3. Compile CUDA Extensions
# (Ensure dependencies are cloned to 3rdparty/)
cd 3rdparty/infllmv2_cuda_impl && python setup.py install && cd ../..
cd 3rdparty/sparse_kernel && python setup.py install && cd ../..
# 4. Install extra deps
uv pip install tilelang flash-linear-attentionQ: CUDA extension compilation failed?
- Ensure CUDA 12+ is installed (
nvcc --version). - Ensure
gcc/g++are available. - If
CXXis set toclang++ -pthread, manuallyexport CXX=g++.
8B-scale trainable sparse attention with hybrid reasoning (2025-09 / 2025-06). MiniCPM4.1-8B · MiniCPM4-8B · ModelScope
Click to expand: highlights, evaluation, inference (HF / vLLM / SGLang / CPM.cu / llama.cpp / Ollama)
MiniCPM 4.1-8B is the first open-source reasoning LLM with trainable sparse attention:
✅ Strong Reasoning Capability: Surpasses similar-sized models on 15 tasks!
✅ Fast Generation: 3x decoding speedup for reasoning
✅ Efficient Architecture: Trainable sparse attention, frequency-ranked speculative decoding
MiniCPM4 and MiniCPM4.1 series are large language models designed for end-side devices, with efficiency optimizations across model architecture, training data, training algorithms, and inference systems.
-
🏗️ Efficient Model Architecture:
- 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 (InfLLM-V2 Training Kernels)
-
🧠 Efficient Learning Algorithms:
- Model Wind Tunnel 2.0 -- Efficient Predictable Scaling: Introduces scaling prediction methods for performance of downstream tasks, enabling more precise model training configuration search
- BitCPM -- Ternary Quantization: Compresses model parameter bit-width to 3 values, achieving 90% model bit-width reduction
- Efficient Training Engineering Optimization: Adopts FP8 low-precision computing technology combined with Multi-token Prediction training strategy
-
📚 High-Quality Training Data:
- 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
- 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
-
⚡ Efficient Inference and Deployment System:
- CPM.cu -- Lightweight and Efficient CUDA Inference Framework: Integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding (Inference Kernels and Framework)
- ArkInfer -- Cross-platform Deployment System: Supports efficient deployment across multiple backend environments, providing flexible cross-platform adaptation capabilities
On two typical end-side chips, Jetson AGX Orin and RTX 4090, MiniCPM4 and MiniCPM4.1 show faster processing speed than similar-size models in long-text processing tasks. As text length increases, the speed gains become more pronounced. On Jetson AGX Orin, compared with Qwen3-8B, MiniCPM4 and MiniCPM4.1 achieve approximately 7x decoding speed improvement.
MiniCPM4.1 achieves 3x decoding speed improvement in reasoning.
MiniCPM4 launches end-side versions with 8B and 0.5B parameter scales, both showing competitive performance in their respective categories.
MiniCPM4.1 launches an 8B end-side version with competitive performance in deep reasoning mode.
MiniCPM4 is pre-trained on 32K long texts and achieves length extension through YaRN. In the 128K needle-in-a-haystack task, MiniCPM4 maintains stable performance. MiniCPM4.1 is pre-trained on 64K long texts and also uses YaRN for length extension, with stable performance on the 128K needle-in-a-haystack task.
MiniCPM 4.1 can be used with the following frameworks: Huggingface Transformers, SGLang, vLLM, and CPM.cu. For inference efficiency, CPM.cu is a good first option.
MiniCPM4/MiniCPM4.1 supports both dense attention inference and sparse attention inference modes, where vLLM and SGLang currently only support dense inference mode. If you want to use sparse inference mode, please use Huggingface Transformers and CPM.cu.
- Dense attention inference: vLLM, SGLang, Huggingface Transformers
- Sparse attention inference: Huggingface Transformers, CPM.cu
MiniCPM4.1 supports hybrid reasoning mode, which can be used in both deep reasoning mode and non-reasoning mode. To enable hybrid reasoning mode. User can set enable_thinking=True in tokenizer.apply_chat_template to enable hybrid reasoning mode, and set enable_thinking=False to enable non-reasoning mode. Similarly, user can directly add /no_think at the end of the query to enable non-reasoning mode. If not add any special token or add /think at the end of the query, the model will enable reasoning mode.
# Enable reasoning mode
prompt_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True
)
# Enable non-reasoning mode
prompt_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False
)- Inference with Dense Attention
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
torch.manual_seed(0)
path = 'openbmb/MiniCPM4.1-8B'
device = "cuda"
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
# User can directly use the chat interface
# responds, history = model.chat(tokenizer, "Write an article about Artificial Intelligence.", temperature=0.7, top_p=0.7)
# print(responds)
# User can also use the generate interface
messages = [
{"role": "user", "content": "Write an article about Artificial Intelligence."},
]
prompt_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
model_inputs = tokenizer([prompt_text], return_tensors="pt").to(device)
model_outputs = model.generate(
**model_inputs,
max_new_tokens=32768,
top_p=0.95,
temperature=0.6
)
output_token_ids = [
model_outputs[i][len(model_inputs[i]):] for i in range(len(model_inputs['input_ids']))
]
responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]
print(responses)- Inference with Sparse Attention This model supports InfLLM v2, a sparse attention mechanism designed for efficient long-sequence inference. It requires the infllmv2_cuda_impl library.
You can install it by running the following command:
git clone -b feature_infer https://github.com/OpenBMB/infllmv2_cuda_impl.git
cd infllmv2_cuda_impl
git submodule update --init --recursive
pip install -e . # or python setup.py install To enable InfLLM v2, you need to add the sparse_config field in config.json:
{
...,
"sparse_config": {
"kernel_size": 32,
"kernel_stride": 16,
"init_blocks": 1,
"block_size": 64,
"window_size": 2048,
"topk": 64,
"use_nope": false,
"dense_len": 8192
}
}These parameters control the behavior of InfLLM v2:
kernel_size(default: 32): The size of semantic kernels.kernel_stride(default: 16): The stride between adjacent kernels.init_blocks(default: 1): The number of initial blocks that every query token attends to. This ensures attention to the beginning of the sequence.block_size(default: 64): The block size for key-value blocks.window_size(default: 2048): The size of the local sliding window.topk(default: 64): The specifies that each token computes attention with only the top-k most relevant key-value blocks.use_nope(default: false): Whether to use the NOPE technique in block selection for improved performance.dense_len(default: 8192): Since Sparse Attention offers limited benefits for short sequences, the model can use standard (dense) attention for shorter texts. The model will use dense attention for sequences with a token length belowdense_lenand switch to sparse attention for sequences exceeding this length. Set this to-1to always use sparse attention regardless of sequence length.
- Long Context Extension MiniCPM4.1 natively supports context lengths of up to 65,536(64k) tokens. For conversations where the total length (including both input and output) significantly exceeds this limit, we recommend using RoPE scaling techniques for effective handling of long texts. By modifying the LongRoPE factor, the model can stably handle context lengths of up to 131,072 tokens.
You can apply the LongRoPE factor modification by modifying the model files. Specifically, in the config.json file, adjust the rope_scaling fields.
{
...,
"rope_scaling": {
"rope_type": "longrope",
"long_factor": [0.9982316082870437, 1.033048153422584, 1.0749920956484724, 1.1255096879436193, 1.1863348602111476, 1.259543828902579, 1.3476188888731149, 1.4535223827776373, 1.5807816745852985, 1.7335856049489526, 1.9168922912975785, 2.1365471404135326, 2.3994084200118646, 2.713475511863602, 3.0880118452194134, 3.533650295140154, 4.062463396503134, 4.687974098908333, 5.425075306704039, 6.289818967956352, 7.29902962722721, 8.6357018163639, 10.210822723989212, 12.053807765671676, 14.193944598909404, 16.65780676784363, 19.463620727694074, 22.628311203524586, 26.150106147261315, 30.02526691405111, 34.23183327975347, 38.73811934094828, 43.502489489729555, 48.47627117965394, 53.61139491762471, 58.857366522037935, 64.16798299215064, 69.51359464319125, 74.86555458220285, 80.21497790341579, 85.55322183307433, 90.89611806932027, 96.26245306514224, 101.68269304046481, 107.18619510219668, 112.82253283014026, 118.63764063163615, 119.88866203644656, 120.9462882391725, 121.837565139014, 122.58663780572562, 123.2147719894291, 123.74049454862576, 124.17980424685767, 124.54641761955492, 124.85202548028222, 125.10654406389756, 125.31835105170659, 125.49450117164764, 125.64091910903052, 125.76256945356558, 125.86360463815589, 125.94749252260765, 126.01712561287873],
"short_factor": [0.9982316082870437, 1.033048153422584, 1.0749920956484724, 1.1255096879436193, 1.1863348602111476, 1.259543828902579, 1.3476188888731149, 1.4535223827776373, 1.5807816745852985, 1.7335856049489526, 1.9168922912975785, 2.1365471404135326, 2.3994084200118646, 2.713475511863602, 3.0880118452194134, 3.533650295140154, 4.062463396503134, 4.687974098908333, 5.425075306704039, 6.289818967956352, 7.29902962722721, 8.6357018163639, 10.210822723989212, 12.053807765671676, 14.193944598909404, 16.65780676784363, 19.463620727694074, 22.628311203524586, 26.150106147261315, 30.02526691405111, 34.23183327975347, 38.73811934094828, 43.502489489729555, 48.47627117965394, 53.61139491762471, 58.857366522037935, 64.16798299215064, 69.51359464319125, 74.86555458220285, 80.21497790341579, 85.55322183307433, 90.89611806932027, 96.26245306514224, 101.68269304046481, 107.18619510219668, 112.82253283014026, 118.63764063163615, 119.88866203644656, 120.9462882391725, 121.837565139014, 122.58663780572562, 123.2147719894291, 123.74049454862576, 124.17980424685767, 124.54641761955492, 124.85202548028222, 125.10654406389756, 125.31835105170659, 125.49450117164764, 125.64091910903052, 125.76256945356558, 125.86360463815589, 125.94749252260765, 126.01712561287873],
"original_max_position_embeddings": 65536
}
}For accelerated inference with speculative decoding using vLLM, follow these steps:
First, download the MiniCPM4.1 draft model:
cd /your_path
git clone https://huggingface.co/openbmb/MiniCPM4.1-8B-Eagle3The EAGLE3 vLLM PR has been submitted. For now, use our repository for installation:
git clone https://github.com/LDLINGLINGLING/vllm.git
cd vllm
pip install -e .Start the vLLM inference server with speculative decoding enabled. Make sure to update the model path in the speculative-config to point to your downloaded MiniCPM4_1-8B-Eagle3-bf16 folder:
VLLM_USE_V1=1 \
vllm serve openbmb/MiniCPM4.1-8B \
--seed 42 \
--trust-remote-code \
--speculative-config '{
"model": "your/path/MiniCPM4_1-8B-Eagle3-bf16",
"num_speculative_tokens": 3,
"method": "eagle3",
"draft_tensor_parallel_size": 1
}'The client usage remains the same for both standard and speculative decoding:
import openai
client = openai.Client(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="openbmb/MiniCPM4.1-8B",
messages=[
{"role": "user", "content": "Write an article about Artificial Intelligence."},
],
temperature=0.6,
max_tokens=32768,
extra_body=dict(add_special_tokens=True), # Ensures special tokens are added for chat template
)
print(response.choices[0].message.content)VLLM_USE_V1=1: Enables vLLM v1 API--speculative-config: JSON configuration for speculative decodingmodel: Path to the draft model for speculationnum_speculative_tokens: Number of speculative tokens (default: 3)method: Speculative decoding method (eagle3)draft_tensor_parallel_size: Tensor parallel size for draft model (default: 1)
--seed: Random seed for reproducibility--trust-remote-code: Allow execution of remote code for custom models
For now, you need to install the latest version of vLLM.
pip install -U vllm \
--pre \
--extra-index-url https://wheels.vllm.ai/nightlyThen you can inference MiniCPM4.1-8B with vLLM:
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
model_name = "openbmb/MiniCPM4.1-8B"
prompt = [{"role": "user", "content": "Write an article about Artificial Intelligence."}]
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
input_text = tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=True)
llm = LLM(
model=model_name,
trust_remote_code=True,
max_num_batched_tokens=65536,
dtype="bfloat16",
gpu_memory_utilization=0.8,
)
sampling_params = SamplingParams(top_p=0.95, temperature=0.6, max_tokens=32768)
outputs = llm.generate(prompts=input_text, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)Also, you can start the inference server by running the following command:
Note: In vLLM's chat API,
add_special_tokensisFalseby default. This means important special tokens—such as the beginning-of-sequence (BOS) token—will not be added automatically. To ensure the input prompt is correctly formatted for the model, you should explicitly setextra_body={"add_special_tokens": True}.
vllm serve openbmb/MiniCPM4.1-8B --trust-remote-codeThen you can use the chat interface by running the following code:
import openai
client = openai.Client(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="openbmb/MiniCPM4.1-8B",
messages=[
{"role": "user", "content": "Write an article about Artificial Intelligence."},
],
temperature=0.6,
max_tokens=32768,
extra_body=dict(add_special_tokens=True), # Ensures special tokens are added for chat template
)
print(response.choices[0].message.content)For accelerated inference with speculative decoding, follow these steps:
First, download the MiniCPM4.1 draft model:
cd /your_path
git clone https://huggingface.co/openbmb/MiniCPM4.1-8B-Eagle3The EAGLE3 adaptation PR has been submitted. For now, use our repository for installation:
git clone https://github.com/LDLINGLINGLING/sglang.git
cd sglang
pip install -e .Start the SGLang server with speculative decoding enabled:
python -m sglang.launch_server \
--model-path "openbmb/MiniCPM4.1-8B" \
--host "127.0.0.1" \
--port 30002 \
--mem-fraction-static 0.9 \
--speculative-algorithm EAGLE3 \
--speculative-draft-model-path "your/path/MiniCPM4_1-8B-Eagle3-bf16" \
--speculative-num-steps 3 \
--speculative-eagle-topk 1 \
--speculative-num-draft-tokens 32 \
--temperature 0.7The client usage remains the same for both standard and speculative decoding:
import openai
client = openai.Client(base_url=f"http://localhost:30002/v1", api_key="None")
response = client.chat.completions.create(
model="openbmb/MiniCPM4.1-8B",
messages=[
{"role": "user", "content": "Write an article about Artificial Intelligence."},
],
temperature=0.6,
max_tokens=32768,
)
print(response.choices[0].message.content)Note: Make sure to update the port number in the client code to match the server port (30002 in the speculative decoding example).
--speculative-algorithm EAGLE3: Enables EAGLE3 speculative decoding--speculative-draft-model-path: Path to the draft model for speculation--speculative-num-steps: Number of speculative steps (default: 3)--speculative-eagle-topk: Top-k parameter for EAGLE (default: 1)--speculative-num-draft-tokens: Number of draft tokens (default: 32)--mem-fraction-static: Memory fraction for static allocation (default: 0.9)
For now, you need to install our forked version of SGLang.
git clone -b openbmb https://github.com/OpenBMB/sglang.git
cd sglang
pip install --upgrade pip
pip install -e "python[all]"You can start the inference server by running the following command:
python -m sglang.launch_server --model openbmb/MiniCPM4.1-8B --trust-remote-code --port 30000 --chat-template chatmlThen you can use the chat interface by running the following command:
import openai
client = openai.Client(base_url=f"http://localhost:30000/v1", api_key="None")
response = client.chat.completions.create(
model="openbmb/MiniCPM4.1-8B",
messages=[
{"role": "user", "content": "Write an article about Artificial Intelligence."},
],
temperature=0.6,
max_tokens=32768,
)
print(response.choices[0].message.content)We recommend using CPM.cu for the inference of MiniCPM4 and MiniCPM4.1. 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 and MiniCPM4.1.
You can install CPM.cu by running the following command:
git clone https://github.com/OpenBMB/CPM.cu.git --recursive
cd CPM.cu
python3 setup.py installYou can run the following command to test the speed of the model.
python3 tests/long_prompt_gen.py # generate prompt.txt
python3 tests/test_generate.py --prompt-file prompt.txtYou can run the following command to infer with EAGLE3 speculative decoding algorithm.
python3 -m cpmcu.cli \
--model-path $BASE_MODEL_PATH \
--draft-model-path $EAGLE3_DRAFT_MODEL_PATH \
--prompt-text "Tell me about Tsinghua University" \
--use-eagle3 trueFor more details about CPM.cu, please refer to the repo of CPM.cu.
We also support inference with llama.cpp and Ollama.
You can download the GGUF format of MiniCPM4.1-8B model from huggingface and run it with llama.cpp for efficient CPU or GPU inference.
# case 1: main-cli
./build/bin/llama-cli -m MiniCPM4.1-8B-Q4_K_M.gguf -p "Write an article about Artificial Intelligence." -n 1500
# case 2: server
## launch server
./build/bin/llama-server -m MiniCPM4.1-8B-Q4_K_M.gguf --host 127.0.0.1 --port 8080 -c 4096 -fa on &
## send request
curl -X POST http://127.0.0.1:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [{"role": "user", "content": "Write an article about Artificial Intelligence."}],
"max_tokens": 1500
}'
Please refer to model hub for model download. After installing ollama package, you can use MiniCPM4.1 with following commands:
ollama run openbmb/minicpm4.1
Quantization (BitCPM4) and MiniCPM4 applications (Survey / MCP / Intel AIPC Client): see
docs/README-legacy.md.
- This repository and MiniCPM models are released under the Apache-2.0 License.
- As a language model, MiniCPM generates content by learning from a vast amount of text.
- However, it does not possess the ability to comprehend or express personal opinions or value judgments.
- Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers.
- Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.
This project is developed by the following institutions:
- Please cite our paper: MiniCPM4 if you find our work valuable.
@article{minicpm4,
title={Minicpm4: Ultra-efficient llms on end devices},
author={MiniCPM, Team},
journal={arXiv preprint arXiv:2506.07900},
year={2025}
}