FAQ
This page collects frequently asked questions about OpenBMB model selection, deployment, VRAM, and community use. When you run into a specific error, it helps to have the model version, runtime environment, hardware configuration, launch command and full logs ready, so the issue can be pinpointed quickly.
Models and Resources
Which model should I start with?
Choose first by task modality and device resources:
- Text generation, code and tool calling: Start with the MiniCPM text series.
- Image and video understanding: Choose the MiniCPM-V multimodal series.
- Real-time interaction across vision, speech and text: See the MiniCPM-o omnimodal series.
- Speech generation and voice cloning: See the VoxCPM speech series.
You can compare the parameter counts, context lengths, capabilities and deployment entry points of different versions in the Full Product Catalog.
Where can I find deployment and fine-tuning tutorials?
Go to the Community Technical Library and browse by category: MiniCPM text, MiniCPM-V, MiniCPM-o, VoxCPM or PilotDeck. The Full Product Catalog also links to the corresponding deployment and fine-tuning guides.
I'd like to try a complete application first. Any recommendations?
You can start with MiniCPM Desk Pet. This project covers system requirements, download and installation, first launch, model management and known limitations, making it a good way to quickly try a local model application.
Deployment and VRAM
How can I tell whether my device can run a given model?
Focus on the model's parameter count, precision or quantization method, context length, and the extra overhead of the inference framework. Start by checking the VRAM estimation formula, model quick-reference table and device recommendations on the VRAM & Performance page.
Simplified estimate: total inference VRAM ≈ model weight VRAM + KV Cache + runtime overhead. Long context, high concurrency and multimodal input all increase actual usage, so leave some headroom.
What should I do first when I run out of VRAM or hit OOM?
- Shorten the context length, or reduce the batch size or concurrency.
- Switch to a quantized version such as INT8, INT4, AWQ, GPTQ or GGUF.
- Choose a model with fewer parameters, or enable CPU / GPU hybrid offloading.
- Close other processes using VRAM, and confirm how much VRAM is actually available.
- If that still doesn't help, record the peak usage, launch parameters and full error before asking for help.
Why can't my device run a model even though it supports very long context?
"Maximum supported context" doesn't mean every device can use it directly. The longer the context, the larger the KV Cache typically gets, and computation time increases too. The actual usable length depends on the model architecture, number of KV heads, precision, concurrency and device memory. We recommend starting with a shorter context and increasing it gradually while watching VRAM and speed.
What effects does quantization have?
Quantization can usually reduce VRAM and disk usage significantly, and makes inference feasible on more devices; the cost may be changes in generation quality, speed or framework compatibility. Prefer quantized versions officially provided for the target model or widely validated by the community, and evaluate them on your own real tasks.
Troubleshooting and Community
What if model downloads are slow or fail midway?
- Check disk space, network connectivity and permissions on the download directory.
- Check whether the repository page offers mirrors, resumable downloads or manual download instructions.
- When downloading manually, make sure the files are complete and the directory structure matches the tutorial's requirements.
- Don't just report "download failed" — include the download method, target repository, file name and the full error message.
What information should I include when asking about installation or launch failures?
| Information | Examples |
|---|---|
| Model and version | Model name, weights or quantized version, code commit or release version |
| Runtime environment | OS, Python, CUDA, PyTorch and inference framework versions |
| Hardware | CPU, GPU model, VRAM and system memory |
| Reproduction steps | Installation method, launch command, sample input and minimal reproduction code |
| Error message | The full log from where the exception starts to where it ends, not just the last line |
Where can I find community projects or get in touch with others?
Browse real projects and technical articles in the Developer Showcase; the "Join the Community" section on the knowledge base home page provides entry points to the MiniCPM, MiniCPM-V, VoxCPM and PilotDeck community groups.
How can I contribute a case study, tutorial or bug fix?
You're welcome to share complete projects, deployment scripts, reproduction notes, benchmarks, VRAM optimizations, failure cases or real-world business experience. Before submitting, please clearly state the applicable model, environment versions, reproduction steps, results and limitations, so the community can verify and reuse your work.
Didn't find an answer?
If the questions above don't cover your situation, feel free to submit an open question. We'll regularly compile and answer them, and common questions may also be added to this knowledge base.