←
OpenBMB2,309 posts
OpenBMB avatar

OpenBMB

@OpenBMB

OpenBMB (Open Lab for Big Model Base) aims to build foundation models and systems towards AGI. Connect with us: discord.gg/7q3ry8Ny8K

◷ Joined February 2022▧ 420 Photos and videos
OpenBMB avatar
OpenBMB@OpenBMB·Oct 27

Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly. 🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results. Two new models: 🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast. 🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench. ⚡ What’s new? Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8). Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute. Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS. Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills. 🔗 GitHub: github.com/OpenBMB/MiniCPM-O… 🤗 Hugging Face: MiniCPM-OCR 1.0:huggingface.co/openbmb/MiniC… MiniCPM-OCR Jev:huggingface.co/openbmb/MiniC… 👉 Try the Demo:platform.modelbest.ai/consol… Use the API:platform.modelbest.ai/consol…

Post attachmentPost attachmentPost attachment
OpenBMB avatar
OpenBMB@OpenBMB·55m

📄 MiniCPM-OCR is landing across the ecosystem. 10 integrations to check out 👇 📚 RAGFlow · PDF parsing github.com/infiniflow/ragflo… 📝 Docling · VLM preset github.com/docling-project/d… 🔗 Haystack · Document converter github.com/deepset-ai/haysta… ⚡ LiteParse · OCR service adapter github.com/run-llama/litepar… 💻 LocalAI · Quantized models + DSpark github.com/mudler/LocalAI/co… 📱 MNN · Model export & inference github.com/alibaba/MNN/pull/… ⚙️ Xinference · Built-in model support github.com/xorbitsai/inferen… 🧩 Dify · PDF & image OCR plugin github.com/langgenius/dify-p… 🧪 lmms-eval · Evaluation wrapper github.com/EvolvingLMMs-Lab/… 📂 kordoc · OcrProvider example github.com/chrisryugj/kordoc… Huge thanks to the maintainers and contributors 💙 Pick your stack and start building! #MiniCPM #OpenBMB #OCR

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·1h

Hybrid architectures: attention and RNN layers compete to learn the same skill 📈 New paper “Competing Learning in Hybrid Architectures” from Tsinghua: • Competing Learning: recurrence learns retrieval first, leaving less learning signal for softmax attention • BLD: freeze recurrent blocks for the first ~10% of pretraining tokens, then unfreeze and train jointly • Better RULER and recall across six KDA hybrids (180M–1B) • At 1B: RULER 32.75 → 36.04; Recall 36.50 → 38.14; LongPPL 7.69 → 6.50 Same architecture, data and token budget. Who learns first matters. 📄 arxiv.org/abs/2610.26830 💻 github.com/thunlp/competing-… #LLM #LongContext #HybridArchitecture #Attention #AI

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·13h

Great to see MiniCPM powering a fully local, privacy-first workflow! 💙 Love how you're combining structured analysis with on-device AI. Keep building! 🙌

materia A cited clinical decision-support engine for naturopathic and integrative practitioners. Blood tests in; cited patterns, supplement/herb options, safety screens and a grounded conversation out. It is not a replacement for a licensed practitioner. It does not issue a final diagnosis or a prescription. Every option is a suggestion for a practitioner to accept, modify or reject. The engine is deterministic. The chat layer runs on CPU with a @OpenBMB 2B GGUF model; no GPU, vLLM or cloud endpoint is required. Privacy first for your clients. I'd love to build this more with those in the industry.

1:00
OpenBMB avatar
OpenBMB@OpenBMB·13h

Love seeing MiniCPM supported in moochy chat! 💙 Fully local AI right in the browser is exciting. Great work! 🚀

As a continuation to this project, I'm currently working on something fun on top of it. Introducing moochy chat: a local AI chat that runs entirely in your browser tab(CPU or GPU), meaning you don't need internet connection to use it after model gets downloaded in the browser tab. got4 is < 54Kb of Web Assembly. The page first load is 135Kb. For some models you can go up to 150 tokens per seconds. No subscription, No API key, 100% free. For now, faster than WebLLM(@mlc_ai_ ), wllama(@ngxson ) and Transformers.js(@xenovacom) on benchmarks. Models shipped right now are : @OpenBMB (MiniCPM), @Alibaba_Qwen, @MistralAI, @AIatMeta (Llama), @deepseek_ai, @huggingface (SmolLM) #LocalLLM #WebGPU #BuildInPublic

2:00
OpenBMB avatar
OpenBMB@OpenBMB·Nov 17

🎬 A new language. A familiar voice. Community developer @kadirrbrtt built VoxaDub, a local-first video dubbing tool that brings VoxCPM2 voice cloning into a multilingual workflow. 🌍 From source video to dubbed output: 📝 Transcribe & align: Transcribe dialogue and align timestamps, with optional speaker diarization to distinguish who speaks when. 🧠 Translate: Local Hy-MT2 translation uses surrounding context and optional terminology. 🎙️ Generate voices: Selected clean audio clips from each speaker guide VoxCPM2 synthesis, helping preserve distinct voice characteristics in the translated dialogue. ⏱️ Fit the timing: Duration-aware translation refinement and bounded audio adjustments help fit speech into the original scene. 🎞️ Rebuild the video: Combine dubbed speech with the original background audio. Optional lip-sync is available as a separate step. 🔒 Offline dubbing can run fully locally after model setup, with optional cloud compute choices. ⚡ VoxaDub also explores live broadcast dubbing, currently experimental. Timing safeguards help improve alignment without promising perfect synchronization. We love seeing our community turn VoxCPM2’s expressive speech generation into tools that help stories reach new audiences. 💙 🛠️ Explore VoxaDub and its multilingual demos: github.com/kadirb4rut/voxadu… 🚀 VoxCPM GitHub: github.com/OpenBMB/VoxCPM 🤗 VoxCPM2 on Hugging Face: huggingface.co/openbmb/VoxCP… #VoxCPM2 #OpenBMB #AIDubbing #OpenSource

1:24
OpenBMB avatar
OpenBMB@OpenBMB·Nov 17

Love seeing MiniCPM-V 4.7 1B explored for computer use! 💙 Great to see the progress from the base model. Looking forward to your next experiments with RL! 🙌

I built an interesting computer use-model at 1B, although nothing crazy. The challenge was to outperform qwen 3.5 4B, and I lost it :D anyways here's HF, still improved from 0/99 base @MiniCPM to 5/99 solved tho. Need to do RL w more compute next time. huggingface.co/banyohavlusu2…

OpenBMB avatar
OpenBMB@OpenBMB·Nov 17

Love seeing MiniCPM5-2B used for a privacy-first, fully local health data workflow! 💙 Turning years of reports into searchable insights is a great real-world use case. Thanks for sharing! 🙌

Sodawhite @sodawhite_dev Nov 16 Article 在 Mac 上管理自己的健康数据,从 8 份体检报告开始 我平时会把体检报告传到手机上的健康助手里,请它帮我看。 拍照时,我会下意识遮一下个人信息,但报告还是从我的实名账号传出去的。这次我想试试,能不能把报告留在自己的 Mac 上,连识别和问答也在本机完成。

OpenBMB avatar
OpenBMB@OpenBMB·Nov 17

Great to see MiniCPM models getting more attention in the local AI community! 💙 Thanks for including them in your roundup! 🙌

Free uncensored AI models 1. MiniCPM-V 4.7 1B Heretic -> criteria: 0.67 GB (Q4), image input -> link: huggingface.co/saidutta69/Mi… 2. Qwen3.8 2B -> criteria: 1.27 GB, 3/100 refusals -> link: huggingface.co/atakhadivi/Qw… 3. MiniCPM5 2B Heretic -> criteria: 1.56 GB, 5/100 refusals -> link: huggingface.co/saidutta69/Mi… 4. LFM2.5 2.6B -> criteria: 1.67 GB, 128K context -> link: huggingface.co/noctrex/LFM2.… 5. Spark X2.6 4B Abliterated -> criteria: 2.60 GB, 1M context -> link: huggingface.co/FrostWolfAI/S… 6. Huihui Qwen3.5 4B Abliterated -> criteria: 2.71 GB, solid general -> link: huggingface.co/mradermacher/… 7. Gemma 4 E2B Heretic ARA -> criteria: 3.43 GB, image + audio -> link: huggingface.co/mradermacher/… 8. LFM2.5 8B-A1B -> criteria: 5.16 GB, MoE -> link: huggingface.co/virtuanista/L… 9. Qwythos 9B 1M Heretic -> criteria: 6.22 GB, 1M context -> link: huggingface.co/llmfan46/Qwyt… 10. Qwen3.6 12B Heretic Uncensored -> criteria: 7.32 GB (Q4), 256K -> link: huggingface.co/larue316/Qwen… nfa dyor save it for don't miss this models btw you can run this models without gpu ^^

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 17

Amazing demo! 🚀 Love seeing MiniCPM5-2B handle real-world, multi-step agent tasks entirely locally. Thanks for showcasing what small models can achieve! 💙

Insane progress for small language models! MiniCPM5-2B is a dense 2B-parameter model by OpenBMB from China that is built for reasoning, coding, and tool use on resource-constrained hardware. The model specifically excels at coding and tool calling, two capabilities central to the shift from on-device LLMs to on-device agents. Instead of only answering prompts, it can use tools, generate code, carry information between steps, and complete multi-step tasks. We ran it 100% locally and connected it to a small investigation agent with one request: > Revenue dropped last week. Investigate what happened, quantify the impact, identify the likely cause, and produce an incident report with supporting evidence. The evidence was spread across orders, traffic, payments, refunds, and deployment logs. The model inspected the files, wrote its own queries, analyzed the intermediate results, and decided what to investigate next. Each tool result informed the next action, so the final report depended on the model maintaining a coherent investigation across the complete trajectory. The recording shows the actual task from beginning to end. It starts with the revenue question, follows the tool calls and supporting evidence, and ends with a quantified diagnosis and incident report. The data, tool execution, and model inference all remained on our machine. These capabilities were optimized through Agentic Pre-training, SFT, and large-scale RL. They do not come entirely from an application-level agent framework. MiniCPM5-2B supports SGLang, vLLM, llama.cpp, Ollama, iOS, Android, and HarmonyOS. OpenBMB has also released the model weights and parts of the training recipes and data resources behind it. A 2B model can now maintain enough state to coordinate tools and complete a useful investigation on local hardware.

0:30
OpenBMB avatar
OpenBMB@OpenBMB·Nov 16

Long-context training: don't stop too early 📈 New empirical scaling law on long-context performance: • Scaing Law on log(LongPPL) ≈ f(N, Ds, Dl): additive power law, extrapolates to held-out scales (R²=0.97) • Constant-Last4 protocol: stable monitoring on one extensible training trajectory • Gains transfer: 32K → 128K, after SFT, and in long-context and agentic tasks 📄 arxiv.org/abs/2610.29352 💻 github.com/thunlp/dont-stop-… 🤗 huggingface.co/collections/E… #LLM #LongContext #ScalingLaw #ContinualPretraining #AI #Agentic

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 16

🚀 MiniCPM-OCR 🤝 vLLM & SGLang Native support for two leading open-source inference frameworks — @vllm_project and @sgl_project — making it easier for developers to integrate MiniCPM-OCR into their existing workflows. ⚡ Powerful OCR, flexible deployment, and a growing open-source ecosystem.

Post attachment

Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly. 🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results. Two new models: 🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast. 🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench. ⚡ What’s new? Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8). Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute. Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS. Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills. 🔗 GitHub: github.com/OpenBMB/MiniCPM-O… 🤗 Hugging Face: MiniCPM-OCR 1.0:huggingface.co/openbmb/MiniC… MiniCPM-OCR Jev:huggingface.co/openbmb/MiniC… 👉 Try the Demo:platform.modelbest.ai/consol… Use the API:platform.modelbest.ai/consol…

Post attachmentPost attachmentPost attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 14

This is awesome, Josh! 🚀 Love how MiniCPM cook makes it easier for AI agents to navigate the MiniCPM family and fine-tuning workflows. Thanks for building this and sharing it with the community! 💙

Want your agent to cook with @OpenBMB MiniCPM models? I've compiled a dsh-plugin that helps you fine tune & work with the whole MiniCPM family model. Open to contributions and feedback - especially to make this adapted to other coding harnesses. This will save your agent a lot of work by having it a plugin instead of searching externally to find model data. github.com/RegeneratusLabs/d…

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 13

📷 A camera captures the scene. MiniCPM helps make sense of it. CamThink’s vision scene-understanding solution connects NeoEyes edge cameras, the NG4500 AI Box, and the NeoMind platform to turn on-site images into actionable information. 🧠 NeoMind supports MiniCPM-V 4.7 1B as a built-in local vision model for image understanding and structured extraction. Here’s the workflow: 📸 1. Capture & transmit: Cameras take snapshots on a schedule or event trigger and send them to NeoMind. ⚡ 2. Trigger the Agent: An incoming image triggers an Agent, with the image added to its analysis context. 💬 3. Understand the scene: The model follows a task prompt to count people, identify safety violations, assess bin-fill levels, or read meters. 🧩 4. Structure & store: Outputs pass Schema validation and become time-series metrics, with the analysis recorded in JSON. 📊 5. Visualize & connect: Dashboards show results and trends, alerts flag anomalies, and Data Push / OpenAPI connect to business systems. CamThink demonstrates construction-site safety patrols, shelf audits, and bin-fill monitoring. Prompts and output fields can be configured for different tasks, with recognition quality checked against on-site images. 🛠️ 🔒 With local inference, images and analysis results stay within the site’s network. Recognition quality depends on prompts, image clarity, and model capability; critical workflows should retain human review. A practical example of MiniCPM’s visual intelligence connecting field observations with everyday operations. 💙 📖 Explore the solution: wiki.camthink.ai/zh-Hans/doc… 🛠️ Deployment & integration guide: wiki.camthink.ai/zh-Hans/doc… 🚀 MiniCPM-V GitHub: github.com/OpenBMB/MiniCPM-V 🤗 Hugging Face: huggingface.co/openbmb/MiniC… #MiniCPM #OpenBMB #EdgeAI #VisualAI

Post attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 13

Thanks for the detailed real-world testing! 💙 Great to see MiniCPM5 2B performing so well on tool calling and long-context tasks. Love seeing the community explore where different small models shine! 🙌

Raon-1 nano follow-up. @Krafton_AI's launch chart was confident, so I put it through my personal-agent tests against the four other small models cited (on my RTX 3080). Short version: Two models came out on top, each in a different way. Full agent tasks, reasoning off (PA Core, out of 300): - @Krafton_AI Raon-1 nano: 163 - MiniCPM5 2B: 150 - LFM3 1.7B: 149 - Gemma 4 E2B: 124 - Qwen 3.5 2B: 114 Individual skills, reasoning off: - @OpenBMB's MiniCPM5 leads on tool calling (99%) and long context (92%) - Raon is solid but lower (86% and 75%), with strong instruction following (93%) So the model that's best at the individual skills wasn't the one that did best on the full agent tasks. Raon-1 nano is the pick for multi-step agent work. It finished all 96 attempts with zero failures, runs at ~187 tok/s, and is ~2GB at Q4. MiniCPM5 2B is the pick for tool-heavy or long-context work, where it was the most accurate small model I tested. With reasoning on, most in this class run out of thinking budget on the agent tasks, so for a responsive local agent I'd run either with reasoning off. (NB: Not a controlled leaderboard: Raon is 3.2B total (~1.6B core) at Q4, and MiniCPM5 ran at Q8. It's the best each managed on my setup. LFM3 also has the known '</think>' tag issue, so those results are omitted) Amazing how far 2B class intelligence has come along. Nice work @Krafton_AI - Let's get this into the upstream llama build! (and shout-out to @OpenBMB also).👏

Post attachmentPost attachment
OpenBMB avatar
OpenBMB@OpenBMB·Nov 12

📸 Snap your meal. Review the estimate. See your daily intake at a glance. Our ambassador @emwstudio built esp32-intake, a food-logging demo that connects MiniCPM-V’s visual understanding with an iOS app and a circular ESP32 display. 🍽️ Here’s the workflow: 🧠 MiniCPM-V 4.7 35B-A3B identifies food in a photo and estimates calories through the ModelBest API. ✅ You review the estimate before adding it to your daily calorie total. Re-estimate or skip it, with a preview of how the meal would affect your daily target. 📊 The app stores today’s calorie total locally and sends the total and daily target over the local Wi-Fi network to an ESP32-S3 AMOLED display. 🟢🟡🔴 A progress ring, calories remaining — or over target — and matching status colors make the day’s intake easy to check. The model runs online via API; the ESP32 handles the display. Calorie estimates are approximate and intended for general reference. We love seeing our ambassadors turn visual intelligence into hands-on experiences that connect software, hardware, and everyday routines. 💙 🛠️ Explore @emwstudio’s project: github.com/emwstudio/esp32-i… 🚀 MiniCPM-V GitHub: github.com/OpenBMB/MiniCPM-V 🤗 Model on Hugging Face: huggingface.co/openbmb/MiniC… #MiniCPM #OpenBMB #ESP32 #OpenSource

Post attachment