OpenBMB

Names, emails, API keys—Our ambassador @povedaai is putting MiniCPM4-0.5B to work detecting sensitive text. 🔐 The project adds a MiniCPM-powered PII recognizer to Microsoft Presidio, helping identify sensitive information in text for downstream anonymization. Here’s what’s under the hood 👇 🧠 A compact foundation: Built on MiniCPM4-0.5B, adapting a 0.5B-parameter model for a specialized token-classification task. 🛠️ Task-specific fine-tuning: Uses a LoRA adapter trained on NVIDIA’s Nemotron-PII dataset, paired with a custom token-classification head. 🔎 55 PII/PHI labels: Covers information such as names, email addresses, phone numbers, financial identifiers, API keys, passwords, and medical record numbers. 🔗 Presidio integration: Maps model predictions into Presidio’s entity format, allowing the recognizer to work within its analysis pipeline and support downstream anonymization workflows. 💻 Local inference: The recognizer loads the model and adapter locally, offering a path for developers exploring sensitive-text processing in their own environments. We love seeing the community turn MiniCPM into practical, specialized tools. This project is an example of how open models, lightweight adaptation, and existing developer frameworks can come together for a concrete application. 🌍 Working on data preparation, text anonymization, or privacy tooling? Try the demo, explore the adapter, and check out the implementation! 🚀 🧩 Model: huggingface.co/luispoveda93/… 📂 Repo: github.com/PovedaAqui/presid… 🔗 Hugging Face: huggingface.co/openbmb/MiniC… #MiniCPM #OpenBMB #Privacy #OpenSource #PII

ForgeKernel Live has wrapped up with a 9.8% speedup over the FlashMLA baseline on DeepSeek V4.1 FlashMLA Prefill, measured on NVIDIA B200. Now, we’re open-sourcing ForgeKernel! 🚀 A quick recap: ⌛️ Around 22 hours of autonomous optimization 🛠️ 32 kernel versions across 42 iteration rounds ✅ Version 13 first beat the baseline within 10 hours 📈 Version 31 reached 1.098× performance: 472 µs vs. 518 µs The official FlashMLA implementation stayed sealed. The agent could run and time its black-box CLI, but could not read its source, documentation, or profiles. We’re releasing: 1️⃣ The optimized DeepSeek V4.1 fused prefill kernel. 2️⃣ Kernels from our ForgeKernel paper, covering attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution. Across 27 operators, ForgeKernel matches or beats 18 of 22 expert-library kernels and outperforms all 7 published kernels from three kernel-generation systems in our evaluation. 3️⃣ The ForgeKernel harness, so you can optimize your own kernels and explore successive generations of improvement. Thank you to everyone who followed the live run. We’re excited to see what the community builds next! 🔗 Live: openbmb.cn/forgekernel/ 💻 Code: github.com/OpenBMB/ForgeKern…

Hold a button. Request landing. Hear an AI tower respond to your simulated flight state. ✈️🎙️ Powered by MiniCPM5-2B + VoxCPM2. Meet an open-source build from our ambassador @emwstudio: an ESP32-S3 cockpit-style attitude indicator with simulated flight data, voice alerts, and push-to-talk conversations with an AI “tower.” 🎙️ 🧠 MiniCPM5-2B gives the tower its responses. Flight-state data, including altitude, speed, heading, pitch, and roll, is fed into the conversation. The tower can respond to the simulated aircraft’s situation and flag inconsistencies between your request and the instrument readings. 🔊 VoxCPM2 brings the tower’s voice to life. Voice cloning from a reference recording, paired with radio-style filtering, background hiss, and a squelch tail, creates that familiar air-to-ground radio feel. 📡 Hold, speak, release. The ESP32 sends audio over Wi-Fi to a Mac-side pipeline. Qwen3-ASR handles local speech recognition, MiniCPM5-2B generates the reply, and VoxCPM2 synthesizes audio that streams back to the device’s speaker. 🛠️ A creative community-built example of model capabilities meeting interactive hardware, with firmware and voice-pipeline code publicly available for others to explore and build on. 💻 Explore Project: github.com/emwstudio/esp32-a… 🤗 Get Models: huggingface.co/openbmb/MiniC… huggingface.co/openbmb/VoxCP… What would you build with MiniCPM + VoxCPM? 👀 #OpenBMB #MiniCPM #VoxCPM #OpenSource #ESP32

That's a challenging case! 💙 Recognition depends on factors like image quality, stamp placement, and handwriting complexity. Here's an example showing how MiniCPM-OCR 1.0 handles a similar document! 👇

Yes! 💙 MiniCPM-OCR 1.0 supports merging tables across pages, helping preserve their structure and column relationships. Here's an example showing how it works! 👇

Great question! 💙 MiniCPM-OCR 1.0 can recognize handwritten corrections and different checkbox states, but doesn't reliably distinguish crossed-out text yet. Here are a couple of examples! 👇

MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B) now run on Apple Silicon, thanks to our partnership with the @neywalabs team! 🚀 On an M5 Max, OCR 1.0 reaches 6,339 tok/s prefill and 220–235 tok/s decode, while Jev scores OCR results in just 62.9 ms median latency. Check out the blog: runnativ.ai/blog/minicpm-ocr…

🎬 MiniCPM-OCR in action — running locally on iPhone Duo and iPad. Parse fast and accurately. Guard the results. All in your hands. 🙌 Handwritten notes, complex formulas, newspapers, poor-quality scans—real documents put OCR to the test. 📄 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. It achieves 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, combining accurate document parsing with efficient inference. 🛡️ MiniCPM-OCR Jev — A 2B decision model that checks OCR accuracy. It approaches Qwen3.5-397B-A17B on JevOCRBench and works with outputs from different OCR parsers. Watch it compare parsed text, formulas, and tables with the original image and assign 0–1 correctness scores—helping you pick better results and identify what needs human review. 🔓 Open weights. Deploy your way. Run with llama.cpp, vLLM, SGLang, or Ollama. Customize with open-source fine-tuning code. Watch it in action. 👇 🔗 GitHub: github.com/OpenBMB/MiniCPM-O… 🤗 Hugging Face: huggingface.co/spaces/openbm… 👉 Try the Demo：platform.modelbest.ai/consol… Use the API：platform.modelbest.ai/consol…

📱 35B parameters. Real-time inference. On a phone. Meet Lookahead Expert Scheduling (LES), the inference system powering MiniCPM-V 4.7’s 35B-A3B MoE model on edge devices. 🚀 On an iPhone 18 Pro Max, the full system reaches 20.51 tokens per second, a 6.07× speedup over loading expert weights on demand. How? Three coordinated optimizations: - 🔮 Scheduling: Predict and prefetch the next layer’s experts while the current layer computes. - 💾 Storage: Combine contiguous expert layouts with mixed-precision quantization to reduce loading overhead. - ⚡ Compute: Use DSpark speculative decoding to verify multiple candidate tokens in one pass. Predictions guide prefetching; the model’s actual router still determines which experts execute. The architecture, expert counts, and routing remain intact. Mixed-precision quantization reduces the weight file by 26.7%, with just a 0.16-point drop in the five-benchmark average. LES shows how better inference systems can bring larger, more capable models to everyday devices. 🌍 👉 LES: github.com/openbmb/llama.cpp… 🔗 GitHub: github.com/OpenBMB/MiniCPM-V #MiniCPM #OnDeviceAI #MoE #EdgeAI