Welcome to the
OpenBMB Community
From local chat and multimodal interaction to agent apps, explore on-device AI with MiniCPM
Where your next build starts
Full Product Catalog
Explore five model series, agents and other products, from on-device capabilities to app building, and find the open-source tools for your idea.
View all productsMiniCPM Text
From everyday chat to code generation, complex reasoning and tool use, turn your device into an assistant that understands and solves problems.
Get startedMiniCPM-V Multimodal
Read images, parse charts and understand video, giving your app the ability to see and opening up new visual experiences.
Get startedMiniCPM-o Omnimodal
Vision, speech and text together: AI that watches, listens and responds at once, for more natural real-time interaction.
Get startedVoxCPM Speech
From natural speech generation to personalized voice cloning, give your app a voice and make every expression more vivid.
Get startedMiniCPM-Robot Embodied AI
From target tracking to robot manipulation, bring intelligence into the physical world and explore what happens when perception meets action.
Get startedPilotDeck Agent
Hand your goals to agents and explore a way of working from task to execution, letting AI take part in real work and turn ideas into results.
Learn moreDeveloper Showcase
View all →ESP32 Radar Tower: AI ATC CalloutsA round-AMOLED tower radar announcing flights in a cloned ATC voiceSelf-Evolving Catgirl: Reasoning via On-Policy Self DistillationNo teacher, no new base model: math reasoning emerges from self-distillationReJev: Decision-Model Post-TrainingLoRA post-training MiniCPM5-2B: 51.1% → 80.5% on a sealed holdoutBMB Agent: a MiniCPM-powered Personal AssistantLetting a 2B on-device model take over schedules and todos, with real-world lessonsMetaCog: Judge-Steered ReasoningAn inference-time loop where a judge scores parallel thought pathsMiniCPM News Desk: Local Daily BriefA local-first daily brief where MiniCPM5-2B picks exact source passagesPilotDeck Art Director: Supervising AI Image GenerationAn Agent that watches the image model work, re-rating and retrying until it passesLater: a Long-term Life Memory AgentA PilotDeck App butler that remembers your life across sessionsBooking Desk: Full-Duplex Voice BookingA no-fine-tuning full-duplex appointment desk driven by a state machine and judgeVoxWeft: On-Device Interpretation on MacVoxCPM2 ported to MLX for fully local streaming speech-to-speech interpretationWebGPU Pi: In-Browser Coding AgentMiniCPM5-2B running a coding agent fully in-browser via WebGPUGander: Talk While It WorksFull-duplex omni agent that delegates long tasks to Codex mid-conversationKahya TTS: Turkish Voice CloningTurkish TTS fine-tuned from VoxCPM2 with reference-based voice cloningMiniCPM5-1B Tool-Use: Offline Function CallsA 1B GGUF function-calling model (SFT + GRPO) that runs offline on CPUVoiceMem: Streaming Dual-Brain Voice MemoryA streaming dual-brain memory giving voice agents low-latency, emotional memoryYouDub: Open-Source Video LocalizationTranscribe, translate and dub videos with sentence-level voice cloningdeku: A Trustworthy Local Task AgentCode plans, MiniCPM5-1B answers briefly or refusesMiniCPM5-1B: LiteRT & Iberian EvalsINT8 LiteRT-LM build plus Spanish, Catalan, Basque and Galician evalsMiniCPM Agent Gateway: Edge-Cloud RoutingLocal MiniCPM classifies each request so sensitive data stays on-deviceToneBridge: Context-Aware Mandarin CoachCorrects Mandarin sentences by context and tone, with pinyin and audioI Had an Agent Fine-tune a Multimodal Model from Scratch68 minutes later it handed over a plan that just runsTyloQuant: Mixed-Precision QuantizationMFQ mixed-precision quantization that compresses MiniCPM-o 4.5MiniCPM5-1B FP8: Faster NVIDIA InferenceFP8 vLLM profile hitting 248 decode tok/s on RTX 3060, 1.5× BF16AUGURY: Reading Soil Health from WeedsMaps weed species to soil-health indicators; phone-deployable, no cloudgmgn-minicpm: Local Crypto Research AgentA read-only GMGN crypto research agent powered by a local 1B modelGrandpa's Bedtime Stories: Photo TalesA WebXR storybook where Grandpa answers voice questions about photosLost & Found Desk: Small-Model Item ReturnMiniCPM-V captions items, MiniCPM5 guides claims, staff confirm every handoffAI Jarvis: Screen-Aware Desktop PetA local Windows desktop assistant that watches your screen and audioCoreAI Model Zoo: MiniCPM on iPhoneCore AI builds of MiniCPM5 and MiniCPM-V 4.6 running natively on iPhone and Macbaby mynah: Morphing Between 9 AccentsVoxCPM-0.5B with one LoRA per accent, blended in weight spaceViva Mais: Local WhatsApp CRM CopilotA travel-agency copilot that turns WhatsApp chats into a local-first CRMTianwen: BaZi & I-Ching ReflectionA 1B model that turns BaZi and I-Ching readings into self-reflectionWanderLust: Mood-Driven Discovery WalksPlans exploratory A-to-B walking routes based on your moodNIGHTWAVE: 1970s Radio Hosted by a 1B ModelA 1970s late-night radio with a 1B-model DJ who takes live calls6ixPulse: Agentic Toronto Rent MapAgents compare Toronto neighborhoods, citing a source for every claimWall Street of AI AgentsFour small-model AI traders roam a pixel office and trade on their ownHackathon Advisor: Idea Map & CoachMaps hackathon ideas and coaches you toward more original projectsLLM Cinema: AI-Directed ASCII FilmsA small model writes and directs ASCII-art films in the browsertiny-llm: A Pure-Rust Inference EngineA from-scratch, CPU-only Rust inference engine for MiniCPM5-1B with a TUIYRobot: Full-Duplex Reachy MiniFull-duplex audio-visual conversation and lifelike motion for Reachy MiniAranya: Seeing Plants, Telling Their StoriesLoRA-tuned MiniCPM-V 4.6 identifies plants and diagnoses their healthTalk to Fengge: A Real-Time Voice ClonePhone-style voice chat with an AI clone that captures both voice and personaBuilding an Online Smart Data System with PilotDeckLessons from building a smart data management system on PilotDeckDocument Field Agent: Batch OCR & ExtractionA workbench for batch document OCR and structured field extractionWhispera: Local Real-Time Voice ChatFully local voice chat on Windows with barge-in, SenseVoice ASR and VoxCPM streaming TTSWhen a Catgirl Learns to Think, on Your PhoneWas this model made for catgirls? An on-device deployment testMaking a 1B Model a Good Math AgentTeXada’s on-device MiniCPM practice and lessonsNekoAudio-80K: a 300 GB Open Catgirl Audio DatasetGive your catgirl a voice: time to train some TTS!VoxCPM CLI: Local TTS via OpenVINOA JSON-first, PyTorch-free VoxCPM2 CLI on OpenVINO, shipped with an agent SkillORG-2: A Workbench for Coding AgentsAn organizational workbench to orchestrate and manage multiple coding agentsEdgeHome HarnessA replayable, evaluable local control loop for a 1B on-device model, built in RustSqueezing a NAS: Hooking Up WorkBuddyRunning local AI on an N150 CPU: real-world uses and lessons learnedPilotDeck Data Analysis: Attendance, Expenses, AssetsThree office scenarios: from raw records to reports and audit findings in one shotPilotDeck Word: Generating and Editing DOCXBatch-generate documents and apply review comments one by one, with tracked editsMaking Slides with PilotDeckFrom a one-line brief to outline, template and slide-by-slide copyReconciliation Workbench: OCR Tie-OutMiniCPM-V reads financial reports; a deterministic engine catches OCR digit errors
Community
View all →MiniCPM Spark Program📢 Open callOngoing programShare real applications and technical experience with the MiniCPM family, building reusable practicesDetails →MiniCPM Global Ambassador Program📢 Open callOngoing programConnecting developers with the core team through tech sharing, community events and model feedbackDetails →
Join the MiniCPM Community
Together with developers worldwide, bringing large models to everyone
Join community events →FAQ
View all →Common questions about deploying and using on-device models.
Which model should I start with?Choose by task and hardware: MiniCPM Text for text and tool use, MiniCPM-V for image and video, MiniCPM-o for real-time interaction, VoxCPM for speech.Can my device run a given model?Look at parameter count, quantization, context length and framework overhead; the VRAM & Performance page has formulas and a quick reference.Out of memory (OOM): what should I try first?Shorten the context, lower batch size or concurrency, switch to an INT8 / INT4 / GGUF build, or use a smaller model.Why can’t my device use the full context length?KV cache grows with context, so usable length depends on architecture, precision, concurrency and memory. Start short and increase gradually.
What does quantization change?It usually cuts memory and disk use substantially, at a possible cost to quality, speed or compatibility; evaluate on your real tasks.Model downloads are slow or keep failingCheck disk space, network and folder permissions, look for mirrors or manual download notes in the repo, and verify file integrity.Install or launch failed: what should I include when asking?Model and version, environment, hardware, reproduction steps, and the full log from the first error to the end.How can I contribute cases, tutorials or fixes?Share projects, deployment scripts, reproductions and lessons learned, stating the model, versions, steps and limitations.
- Which model should I start with?Choose by task and hardware: MiniCPM Text for text and tool use, MiniCPM-V for image and video, MiniCPM-o for real-time interaction, VoxCPM for speech.
- Can my device run a given model?Look at parameter count, quantization, context length and framework overhead; the VRAM & Performance page has formulas and a quick reference.
- Out of memory (OOM): what should I try first?Shorten the context, lower batch size or concurrency, switch to an INT8 / INT4 / GGUF build, or use a smaller model.
- Why can’t my device use the full context length?KV cache grows with context, so usable length depends on architecture, precision, concurrency and memory. Start short and increase gradually.
- What does quantization change?It usually cuts memory and disk use substantially, at a possible cost to quality, speed or compatibility; evaluate on your real tasks.
- Model downloads are slow or keep failingCheck disk space, network and folder permissions, look for mirrors or manual download notes in the repo, and verify file integrity.
- Install or launch failed: what should I include when asking?Model and version, environment, hardware, reproduction steps, and the full log from the first error to the end.
- How can I contribute cases, tutorials or fixes?Share projects, deployment scripts, reproductions and lessons learned, stating the model, versions, steps and limitations.
Didn’t find your answer? Ask in AMA and we’ll answer regularly.Ask in AMA →