Deploy MiniCPM5-1B with 🤗 Transformers

MiniCPM5-1B is a standard LlamaForCausalLM, so it loads directly via AutoModelForCausalLM — no custom modeling code, no trust_remote_code.

Install

pip install -U "transformers>=5.6,<6" "torch>=2.11" accelerate     # latest (CUDA 13.x driver hosts)
# pip install -U "transformers==4.57.3" "torch==2.7.1" accelerate  # fallback for CUDA 12.x driver hosts

GPU inference (bfloat16)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "openbmb/MiniCPM5-1B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()

messages = [{"role": "user", "content": "用一句话解释什么是 GQA。"}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=True,        # set False for fast / no-think mode
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=0.9,
        top_p=0.95,
    )
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

CPU-only inference

The whole 1.08B-param model is < 4.5 GB in fp32, so it can run on CPU only (laptops, CI machines, no-GPU sanity checks):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "openbmb/MiniCPM5-1B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float32,   # bf16 also works on AVX-512 BF16 / AMX hosts
    device_map="cpu",
).eval()

messages = [{"role": "user", "content": "用一句话解释什么是 GQA。"}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,        # nothink is recommended for CPU latency
    return_tensors="pt",
)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        top_p=0.95,
    )
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Generation defaults

Mode	enable_thinking	temperature	top_p	When to use
Think	True	0.9	0.95	hard reasoning, math, code, multi-step
No-think	False	0.7	0.95	fast assistant, latency-bound

generation_config.json is tuned for think mode by default.

LoRA inference (PEFT)

from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(
    "openbmb/MiniCPM5-1B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "<your_lora_dir>").eval()

Adapters trained against this base load directly with no surgery.