Full Product Catalog
This page collects OpenBMB's currently open-source text models, multimodal models, omnimodal models, embodied AI models, and speech models, along with related products and application projects. Click a model name to open its page in this knowledge base for an overview, deployment docs, fine-tuning guides, and tutorials. Models or products without a corresponding page link to external resources.
Text Models (MiniCPM Text Series)
| Model | Parameters | Context | Highlights | Deployment Guide | Fine-tuning Guide |
|---|---|---|---|---|---|
| MiniCPM5-2B | 2B | 128K | • Reaches SOTA among open-source models of its size, with overall performance competitive with 4B-class models • Stands out on code, math, long context, tool calling, and agent tasks • Built for on-device, local deployment, and resource-constrained scenarios, balancing model capability with deployment cost | ⏭️ | 🔧 |
| MiniCPM5-1B | 1B | 128K | • Reaches SOTA among open-source models of its size • Balances reasoning, tool calling, and code generation • Supports hybrid reasoning: Thinking / Non-thinking | ⏭️ | 🔧 |
| MiniCPM-SALA | 9B | 1M | • Hybrid sparse attention + linear attention architecture • Supports ultra-long contexts of up to a million tokens • Built for long-text and complex-context tasks | ⏭️ | 🔧 |
| MiniCPM4.1-8B | 8B | 128K | • 8B efficient on-device reasoning model • Supports hybrid reasoning: Thinking / Non-thinking • Introduces trainable sparse attention for faster inference | ⏭️ | 🔧 |
| MiniCPM4-8B | 8B | 128K | • Improves inference efficiency with InfLLM v2 sparse attention • Balances long-context processing with local deployment performance | ⏭️ | 🔧 |
| MiniCPM4-0.5B | 0.5B | 32K | • Ultra-lightweight on-device model • Suited to low-resource local deployment and on-device demos • Balances a small parameter count, long context, and inference efficiency | ⏭️ | 🔧 |
Besides the general-purpose models, we also offer specialized versions for different use cases, for example:
- MiniCPM4-Survey: focused on understanding, summarizing, and structurally analyzing long documents such as survey questionnaires, interview transcripts, and research reports;
- MiniCPM4-MCP: covers 16 categories of MCP Servers, supporting single-tool and cross-tool multi-step calls If you're interested in these directions, you're welcome to learn more and try them out on Hugging Face.
Multimodal Models (MiniCPM-V Series)
| Model | Parameters | Supported modalities | Highlights | Deployment Guide | Fine-tuning Guide |
|---|---|---|---|---|---|
| MiniCPM-V 4.7 35B-A3B | 35.2B total / ~3.5B active | Image, video, text | • First MoE model in the series • LES lookahead expert scheduling and on-device inference • 4x / 16x visual token compression | ⏭️ | 🔧 |
| MiniCPM-V 4.7 1B | 1.55B | Image, video, text | • Lightweight dense model with upgraded core capabilities • Single-image, multi-image and video understanding • On-device deployment on iOS / Android / HarmonyOS | ⏭️ | 🔧 |
| MiniCPM-V 4.6 | 1.3B | Image, video, text | • Early ViT compression • 50%+ FLOPs reduction • On-device deployment on iOS / Android / HarmonyOS | ⏭️ | 🔧 |
| MiniCPM-V 4.6 Thinking | 1.3B | Image, video, text | Deep reasoning variant | ⏭️ | 🔧 |
| MiniCPM-V 4.5 | 8B | Image, video, text | • Surpasses GPT-4o-latest • Supports 30+ languages and 1.8MP images | ⏭️ | - |
| MiniCPM-V 4.0 | 4.1B | Image, video, text | • Surpasses GPT-4.1-mini • Multi-image and video understanding | 🔧 | 🔧 |
Omnimodal Models (MiniCPM-o Series)
| Model | Parameters | Modalities | Highlights | Deployment Guide | Fine-tuning Guide |
|---|---|---|---|---|---|
| MiniCPM-o 4.5 | 9B | Vision, speech, text | • Full-duplex real-time interaction • Sees, listens, and speaks at the same time • Supports proactive interaction • Overall capability approaching Gemini 2.5 Flash | ⏭️ | - |
| MiniCPM-o 2.6 | 8B | Vision, speech, text | • Real-time Chinese-English voice conversation • Continuous audio-video stream interaction • Supports voice cloning • Strong vision and OCR capabilities • Efficient on-device inference | - | 🔧 |
Embodied AI Models (MiniCPM-Robot Series)
| Model | Parameters | Task type | Highlights | Deployment Guide | Fine-tuning Guide | Evaluation & Reproduction |
|---|---|---|---|---|---|---|
| MiniCPM-RobotManip | 1.5B | Robot manipulation VLA | • One set of weights covers many manipulation tasks • Supports streaming visual context • Supports multi-view input and action sequence output | ⏭️ | - | 📊 |
| MiniCPM-RobotTrack | 0.9B | Embodied object tracking | • Static, dynamic, and adversarial target tracking • On-device natural-language tracking • Supports real-robot deployment on Unitree Go2 EDU | ⏭️ | 🔧 | 📊 |
Speech Models (VoxCPM Series)
| Model | Parameters | Languages | Highlights | Deployment Guide | Fine-tuning Guide |
|---|---|---|---|---|---|
| VoxCPM2 | 2B | 30+ languages 9 major dialects | • Multilingual TTS, covering eight major Southeast Asian languages • Voice design, controllable cloning, and high-fidelity ultimate cloning • 48kHz streaming output | ⏭️ | 🔧 |
| VoxCPM1.5 | 0.8B | Chinese / English | • Chinese-English bilingual TTS • High-quality timbre (voice) cloning • 44.1kHz output • Supports SFT / LoRA | ⏭️ | ⏭️ |
| VoxCPM-0.5B | 0.5B | Chinese / English | • Lightweight Chinese-English bilingual TTS • Context-aware generation • Zero-shot timbre (voice) cloning • 16kHz output | ⏭️ | ⏭️ |
Products & Agents
| Product | Base model | Type | Highlights | Learn more |
|---|---|---|---|---|
| PilotDeck | — | AI agent platform | Task-oriented AI agent productivity platform | 🔀 |
| MiniCPM Desk Pet | MiniCPM5 | Desktop app | Local desktop pet | 🔀 |
| MiniCPM-V Apps | MiniCPM-V | Mobile app | Fully offline multimodal chat on iOS/Android/HarmonyOS | 🔀 |