OpenBMB

🚀 A new way to refine LLM training data is here! Scaling LLMs is no longer just about collecting more data — it’s about making every token count. Introducing UltraX, our framework for large-scale pre-training data refinement. Instead of rewriting data end-to-end, UltraX uses structured editing functions to precisely improve training data quality. What’s inside? 👇 ✅ Function-calling refinement Fine-grained editing with insertion, deletion, and modification. ✅ Reliable data transformation LAM + DCR pipelines generate high-quality supervision for scalable refinement. ✅ Better data efficiency Achieves the best average performance across five corpora in 1B-model pre-training experiments. No gatekeeping — explore how better data can help build better models. 📄Paper: arxiv.org/abs/2607.08646 💻 GitHub: github.com/openbmb/UltraX 🤗 Hugging Face: huggingface.co/datasets/openb… Modelscope: modelscope.cn/datasets/OpenB… modelscope.cn/models/OpenBMB…

@aiseomastery Absolutely. Instead of replacing original data with rewritten versions, structured editing enables fine-grained improvements while preserving valuable information from the original dataset.😍

@OpenBMB Structured editing instead of rewriting whole datasets is a much more scalable way to refine training data.

🚀 A new way to refine LLM training data is here! Scaling LLMs is no longer just about collecting more data — it’s about making every token count. Introducing UltraX, our framework for large-scale pre-training data refinement. Instead of rewriting data end-to-end, UltraX uses structured editing functions to precisely improve training data quality. What’s inside? 👇 ✅ Function-calling refinement Fine-grained editing with insertion, deletion, and modification. ✅ Reliable data transformation LAM + DCR pipelines generate high-quality supervision for scalable refinement. ✅ Better data efficiency Achieves the best average performance across five corpora in 1B-model pre-training experiments. No gatekeeping — explore how better data can help build better models. 📄Paper: arxiv.org/abs/2607.08646 💻 GitHub: github.com/openbmb/UltraX 🤗 Hugging Face: huggingface.co/datasets/openb… Modelscope: modelscope.cn/datasets/OpenB… modelscope.cn/models/OpenBMB…

Great question! 😊 UltraX is a framework for improving LLM training data. It helps developers and researchers refine large-scale datasets by automatically identifying and applying precise edits — adding missing information, removing low-quality content, and improving text quality.

@OpenBMB What is for?

Built on MiniCPM-V 4.6, Aranya features two specialized fine-tuned models: ✨ Plant Species Identification → Recognizing plants from images with common and scientific names 🔍 Plant Health & Disease Analysis → Identifying visible signs of disease or stress and suggesting appropriate care steps 🤗MiniCPM-V 4.6: huggingface.co/openbmb/MiniCP… Fine-tuned on more than 310K multimodal plant-care conversations, these models demonstrate how a compact general-purpose VLM can become a capable domain specialist. A great example of adapting open-source vision models to solve everyday problems. 🚀 Congratulations! 🎉

Many plant owners face common challenges: 🌱 “What plant is this?” 🔍 “Why are my plants unhealthy?” 💡 “How should I take care of it?” Aranya makes plant care easier by turning a simple photo into personalized guidance: 🌿 Identifies plant species from images 🔍 Detects potential plant diseases 💡 Provides AI-powered care suggestions 🎙️ Delivers results through visual understanding and voice narration 🌳 Creates an engaging jungle-themed plant exploration experience

🏆 Build Small Hackathon Winner—OpenBMB Award: 🥈 Backyard AI Track Congratulations to @getsounth for winning the Backyard AI Track ($1,500 prize) with Aranya: A WildKeeper's Adventure 🎉 Aranya is an AI plant assistant that makes plant care easier by helping users identify plants, understand potential health issues, and receive personalized care guidance powered by MiniCPM-V 4.6👇

@ValencianaAbel Thanks for sharing MiniCPM-SALA’s performance! 🚀 We’d also love to see what developers can build with it. If you create any projects or demos with MiniCPM, we’d be happy to feature them through our official channels.

HUMANEVAL HAS A SURPRISING LEADERBOARD HumanEval is the coding benchmark that asks a model to write Python functions which pass hidden unit tests. CURRENT TOP THREE 1. MiniCPM-SALA: 95.1% 2. Kimi K2 0905: 94.5% 3. Claude 3.5 Sonnet: 93.7% Huge win for teams willing to route work instead of paying flagship prices by habit. My take: when results compress this tightly, verification and task-specific evaluation matter more than the logo on the model picker.

@notcalmerecentl @iamsorrybutno @tugmirk @mertcobanov Thanks for asking! 😊 VoxCPM focuses on high-quality voice generation and voice cloning capabilities. We’d love for you to try it out and compare it with other models. Let us know your feedback!

@iamsorrybutno @tugmirk @mertcobanov voxcpm veya supertronic 3?

eger herdr'i denemediyseniz muhakkak bakın. uc gundur test ediyorum ve harika. temelinde tmux for agents fakat cok cok daha iyi. gelistiren turk bir dostumuz hali hazirda githubda 10k stari gecti ve tum yabanci youtube kanallari bundan bahsediyor.

@vintcessun 感谢分享使用体验！🙌 多项目管理、记忆隔离以及模型调度确实是 AI Agent 工作流中的重要挑战，很高兴看到 PilotDeck 的这些设计能够帮助开发者提升使用体验。期待看到你后续更多探索！🚀

终于有个Agent平台把多项目记忆混乱的问题当正事解决了。说白了，PilotDeck用WorkSpace做隔离，每个项目独立文件系统、记忆和技能，A项目的记忆不会泄露到B项目。更离谱的是记忆是白盒的，哪条错了能直接编辑，不用推倒重来。加上Smart Routing自动根据难度分配模型，简单任务走便宜模型，复杂任务走旗舰，实测能省约70%的token成本。下班前把任务丢进去，它自己会持续干活然后落地文件。的确比自己在胶水上搭要清爽太多。 github.com/OpenBMB/PilotD…

@Ryrenz 感谢分享呀😍🫶

🗣️ 炸裂！这个开源项目免费文字转无 AI 味人声，还能克隆任何人的嗓音，并且用文字调整音色！ GitHub 狂揽 30K star，出自面壁智能 OpenBMB，VoxCPM 之前拿过 GitHub 和 HuggingFace 双榜第一。 做有声书、配音、口播视频的人最清楚，市面上的 TTS 要么一股机器味、要么一开口就露馅。 VoxCPM2 是个 2B 的语音模型，用了两百多万小时语音训练，能直接生成连续语音、读出来自然又带情绪。 它有两大牛逼之处： 一是只给一句话的人声样本就能克隆出这个人的音色，连节奏和情绪都还原； 二是你不用录音，光用文字描述「中年男性、低沉、慢一点」它就能凭空造出一个全新的声音。 支持 30 种语言，连四川话、粤语、上海话这些方言都能说，输出还是 48kHz 录音棚级音质，而且 Apache 协议可商用。 部署也简单，pip install voxcpm 装完就能合成。 配音门槛被它一下拉到地板。 GitHub：github.com/OpenBMB/VoxCPM

@YakuzaDaddy @openclaw @NousResearch Thanks for trying VoxCPM and sharing your experience! 🙌 It's exciting to see VoxCPM being used to bring more personalized voices to AI agents. Looking forward to seeing what you build with it next! 🚀

I was paying for eleven labs so my @openclaw and @NousResearch agents would have a custom voice Until I found this free Voice cloner VoxCPM: github.com/OpenBMB/VoxCPM The audio quality and voice cloner is top tier! Check out my agent Makimas sexy voice 😌

@Honcia13 感谢老师推荐呀🫶

真就离谱了，这玩意儿是真神！ GitHub趋势第一、星标直接干到 2万+ 开源语音合成AI VoxCPM2，海外已经彻底炸锅。 生成的声音你根本听不出来是AI！ 1、打一行字「沉稳的30岁女声」就能直接出 2、扔一小段录音，连口癖、语速、情感都能完美复刻 3、48kHz录音棚级音质，Apache 2.0协议，商用随便用 最离谱的是——全部免费！ github.com/OpenBMB/VoxCPM 配音党、短视频创作者、数字人玩家，直接冲就完事了 #VoxCPM #AI语音合成 #开源神器 #AI配音 #短视频工具

@DAIEvolutionHub Thanks for sharing VoxCPM😍

Someone on GitHub just shared a massive list of free projects that are ridiculously good. Many of them can already replace software people are paying monthly subscriptions for. 1. TradingAgents AI-powered multi-agent quantitative trading framework github.com/TauricResearch… 2. LibreChat A unified interface for ChatGPT, Claude, Gemini, and other AI models github.com/danny-avila/Li… 3. HyperFrames HeyGen's open-source video generation engine github.com/heygen-com/Hyp… 4. Fincept Terminal An open-source alternative to the Bloomberg Terminal github.com/Fincept-Corpor… 5. MoneyPrinterTurbo AI that generates short-form videos with a single click github.com/harry0703/Mone… 6. Agentic Inbox Cloudflare's open-source AI email assistant github.com/cloudflare/age… 7. VoxCPM AI voice cloning platform github.com/OpenBMB/VoxCPM 8. Flowsint Open-source OSINT intelligence analysis tool github.com/reconurge/flow… 9. agent-skills A library of coding skills for Claude github.com/addyosmani/age… 10. Nango Open-source API integration platform github.com/NangoHQ/nango These aren't toy projects. A lot of the software you're still paying monthly for already has open-source alternatives built by developers on GitHub. Some of the most powerful tools on the internet aren't being advertised. They're hidden in GitHub repositories.

@zast57 Amazing work! 🎉 Turning video translation and dubbing into a fully local workflow is a very practical application of AI. We're excited to see VoxCPM 2 being used in creative tools like ZastTranslate. 🚀

🚀 Say hello to local AI video translation and voice cloning! ZastTranslate Beta 1.06 is officially live! 🎬 If you are looking for an open source video translator that runs 100% locally on your computer with zero API keys and no subscriptions, this update is for you and only on Pinokio @cocktailpeanut . We just released a major update to the Gradio web interface, introducing powerful tools to make your automated dubbing workflow smoother: ✨ Interactive Timeline Segment Editor – Take full control of your project by adjusting segment start and end times or editing dialogue text on the fly. 📺 Split-Screen Subtitle Preview – A persistent preview column on the left displays real-time original or translated subtitle overlays synchronized with the video player. 🎯 Click-to-Seek Navigation – Click any dialogue row to instantly jump the media player playhead to the start of that segment. 🔄 Single-Segment Audio Regeneration – Fine-tune and re-synthesize specific lines instantly, leveraging fast local voice cloning without reprocessing the entire video. How the local AI pipeline works under the hood: 1️⃣ Vocal separation using Demucs isolates speech from background music. 2️⃣ WhisperX transcription creates precise word-level timestamps. 3️⃣ Multi-backend translation (Qwen2.5/3.5 or EuroLLM) automatically translates and fits text length to match the video's original timing. 4️⃣ Smart dubbing synthesizes natural-sounding speech, cloning the original speaker's voice in 30+ languages using VoxCPM 2. 📦 Ready to translate, dub, and export? Getting started is incredibly easy! Thanks to the Pinokio launcher, you can set up all dependencies, PyTorch CUDA packages, and models in a single click. It's optimized for NVIDIA GPUs to run smoothly without out-of-memory crashes. 🌍 Supports 30+ languages including English, French, Spanish, German, Japanese, Chinese, Italian, Portuguese, Korean, Russion and more! Try ZastTranslate today and take control of your local translation and speech synthesis!

Thanks for sharing your thoughts! 🙌 We totally agree that open-source voice models are improving rapidly, and different models may shine in different scenarios. We’d love to learn more about your experience — what aspects make Vibevoice work better for conversation compared with VoxCPM? (e.g. naturalness, latency, turn-taking, emotional expression, etc.) Your feedback would be really valuable for us! 🚀

It's actually insane how open source voice models have reached s decent level for a while. VoxCPM and Vibevoice are impressive. Playing around with them is surprisingly good. Although for conversation, Vibevoice is way better than VoxCPM

@Abobsterina Thanks for including our model VoxCPM😍😊

MOST PEOPLE ARE PAYING MONTHLY FOR SOFTWARE THAT ALREADY HAS A FREE OPEN-SOURCE TWIN SITTING ON GITHUB. Ten worth knowing right now. 1.TradingAgents — multi-agent quant trading framework where AI analysts, researchers, and risk managers debate before a trade. github.com/TauricResearch… 2.LibreChat — one interface for ChatGPT, Claude, Gemini, and local models. The multi-model chat you were paying three subscriptions for. github.com/danny-avila/Li… 3.HyperFrames — HeyGen’s open-source video engine. Write HTML, render MP4. Built for agents, ships with 20 skills. github.com/heygen-com/hyp… 4.Fincept Terminal — an open Bloomberg Terminal with AI analyst agents. Bloomberg wants $24K a year. This wants nothing. github.com/Fincept-Corpor… 5.MoneyPrinterTurbo — keyword in, finished short video out: script, footage, captions, music. github.com/harry0703/Mone… 6.Agentic Inbox — Cloudflare’s self-hosted email client with an AI agent. Your mail never leaves your account. github.com/cloudflare/age… 7.VoxCPM — voice cloning from a few seconds of audio, multilingual, tunable by text description. github.com/OpenBMB/VoxCPM 8.Flowsint — graph-based OSINT investigations. Domain in, connected IPs, emails, and accounts out. Everything stays local. github.com/reconurge/flow… 9.agent-skills — Addy Osmani packaged 23 production workflows into Claude Code skills. Fifteen years of Google engineering, one install command. github.com/addyosmani/age… 10.Nango — open-source API integration platform. Hundreds of connectors you were about to build by hand. github.com/NangoHQ/nango The detail most people miss: these are not toy demos. Several are backed by real companies open-sourcing their core as distribution. Stop bookmarking AI tool websites. The serious stuff ships as a repo.

Great point! 👏 Real-world conversations are often ambiguous, and turning them into reliable structured data requires much more than simple extraction. This is exactly why understanding the business context and continuously refining the system with real user feedback matters for practical AI applications.

'we're good for Thursday' can confirm a booking, postpone it, or reference a different trip entirely. same contract as calendar dedup: parseable structure, ground truth in implicit context. built this in Basil for meeting transcripts and had to treat every extracted item as a hypothesis.

🏆 Build Small Hackathon Winner—OpenBMB Award: 🥇Viva Mais AI Congratulations to @amphetamarina for winning the Backyard AI Track ($2,500 prize) with Viva Mais AI! 🎉 More than a CRM, it’s a "digital brain" for small travel agencies. Here’s how it leverages open-source small models to solve real-world business problems. 👇

Behind Viva Mais AI: 🧠 Multimodal document understanding MiniCPM-V 4.6 helps analyze and extract information from travel documents, tickets, and receipts. ⚡ Lightweight local AI processing MiniCPM5-1B enables efficient information extraction and Q&A on local devices. Powered by: MiniCPM-V 4.6： 🤗 Hugging Face: huggingface.co/openbmb/MiniCP… ⚡ MiniCPM5-1B: 🤗 Hugging Face: huggingface.co/openbmb/MiniCP… Built with a clean architecture and custom HTMX frontend, Viva Mais AI shows how small open-source models can power practical business solutions. 🌍 Congratulations to @amphetamarina and all the builders who participated in Build Small Hackathon! 🎉

Small travel agencies often struggle with fragmented information—itineraries, payment records, and travel documents scattered across WhatsApp conversations. Viva Mais AI transforms this chaos into a structured, local-first CRM: ✈️ Automatically analyzes travel conversations 📄 Extracts key information from tickets, payments, and documents 🤖 Provides AI-powered summaries and Q&A 🔒 All processing stays on‑device—your clients’ privacy is never compromised.