Can you run it?
This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Install & run
Pull the weights from the Hub and pin the revision. The card's three routes (ev-card):
pip install "vllm>=0.21" && vllm serve openbmb/MiniCPM5-1B --port 8000
pip install "sglang[srt]>=0.5.12" && python -m sglang.launch_server --model-path openbmb/MiniCPM5-1B --port 30000
pip install -U "transformers>=5.6" accelerate torch # then AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-1B")
For laptops and phones use the official GGUF with llama.cpp (the GGUF card documents llama serve -hf openbmb/MiniCPM5-1B-GGUF, which downloads Q4_K_M) or the official MLX 4-bit on Apple Silicon. The version floors are recent; older vLLM and transformers releases may not load it. Verify the LFS sha256 against the repository listing (ev-repo-files).
Hardware & VRAM requirements
The publisher gives file sizes, not device benchmarks (ev-repo-files, ev-github): BF16 safetensors 2.16 GB; GGUF F16 2.17 GB, Q8_0 1.15 GB, Q4_K_M 0.69 GB (the cookbook rounds this to 657 MB); MLX 4-bit 0.61 GB. The KV cache is small because there are only 2 key-value heads. By arithmetic from config.json (24 layers x 2 heads x 128 head dim x 2 tensors x 2 bytes) it is 24,576 bytes per token, about 3.2 GB at the full 131,072 tokens in 16-bit. On a small device cap the context; the cookbook's llama-server example uses -c 8192. No RAM, speed or battery figures for named phones or boards are published, so measure on your target.
Serving stacks
The model is a standard LlamaForCausalLM, and the card says "no custom kernels, no model-code fork" are needed (ev-card). The card and repository provide cookbooks for Transformers, vLLM, SGLang (recommended for tool calling), llama.cpp, Ollama, LM Studio, MLX, ArcLight, LiteRT-LM (a .litertlm runtime for Android, iOS, desktop and IoT) and vLLM Ascend (ev-github). The card also lists FlagOS multi-chip builds for Nvidia, Hygon, Metax, Iluvatar, Zhenwu, Mthreads, Kunlunxin, Ascend and ARM-v9, published under the separate FlagRelease org. The ollama.com library page for minicpm5 does not exist (404); the cookbook uses a local Modelfile (ollama create minicpm5-1b -f Modelfile).
Safe-deployment controls & Deployment Ceiling
Deployment Ceiling: T1 (internal assistant), conditional. Humans read the output, and the deployer writes the acceptable-use policy, because the publisher provides neither a policy nor a safety evaluation (ev-card). T2 and above need controls the deployer must build:
- Use the final MiniCPM5-1B checkpoint and not Base for any assistant. SFT-only is a research stage.
- Run safety and refusal tests in the languages served, and add input and output guardrails or a separate classifier. None is released with the model.
- Treat tool calls as untrusted model output: parse with SGLang's
minicpm5parser or the repo's tool parser, validate arguments and allow-list actions. - Pin the revision and verify the sha256. Avoid unreviewed community conversions, some of which are named "Uncensored".
- Test long-context behaviour, because no long-context result is published for the 1B.
Available quantizations
Official: GGUF in F16, Q8_0 and Q4_K_M, and an MLX 4-bit conversion (ev-repo-files). The repository also ships AWQ, GPTQ and bitsandbytes quantisation scripts with an evaluation script (ev-github). The later 2B release adds official GPTQ and speculative-decoding drafts; the 1B has none of those from the publisher. The card gives no quality-loss numbers for the 1B quantisations beyond the cookbook description of Q8_0 as "very small quality drop vs F16". Many community quantisations exist (MLX, AWQ, NVFP4, ONNX, MNN, LiteRT-LM); they are not publisher-verified.
Fine-tuning & adaptation
Apache-2.0 allows fine-tuning and redistribution. The Base and SFT-only checkpoints are published alongside the final one. Cookbooks and agent skills cover TRL with PEFT (LoRA and SFT), LLaMA-Factory, ms-swift, unsloth and xtuner; the repository also has finetune scripts including an MLX trainer (ev-card, ev-github). The desk-pet demo uses LoRA persona switching. The public pretraining datasets allow continued pre-training on the published mix. The SFT dataset needs a Hub login and gate acceptance (ev-sft-dataset). RL and OPD are described but their pipelines are not released.
API / OpenAI-compatible integration
vllm serve and the SGLang server both expose /v1/chat/completions. For tool calling the card says SGLang is recommended: start it with --tool-call-parser minicpm5 (or auto) and the model's XML-style calls come back as OpenAI-compatible tool_calls (ev-card). The card gives mode-specific sampling: temperature 0.9 and top_p 0.95 for Think, 0.7 and 0.95 for No Think.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
ReliabilityIs it reliable and good enough for the job?
WeakOperational support is broad (4), but there is no safety evaluation, policy or guard model (safety 1), so the model is unsafe to deploy as presented until the deployer supplies that layer.
Doesn't extract your dataDoes running it keep your knowledge and data yours?
StrongSelf-hosted weights run on the deployer's own device and the Apache-2.0 licence claws back no rights; no phone-home behaviour is documented.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.