OpenBMB @OpenBMB

Dostlar Kahya-TTS yayında 🎉 @OpenBMB 'in harika modeli VoxCPM2'i 100 saate yakın Türkçe metin-konuşma sentezli doğal türkçe ses datasetimle finetune ederek, Türkçe özellinde geliştirmeler sağladım. Uyarım topluluğumuza yararı olur. İyi kullanımlar dilerim ❤️

@papersdatacode Appreciate the write-up! Great to see error recovery called out alongside tool use. Curious to hear how the dataset works for folks fine-tuning their own agents.

@tysyrrr That was our favorite part too — not just changing its answer, but writing a verifier to check it. A really nice example of what small models can do locally. 🙌

Love this. 🙌 MiniCPM5-2B was built for exactly this: on-device, offline, download from Hugging Face and run. 2.5B, native 128K, hybrid Think / No-Think in one checkpoint. First calling the puzzle impossible, then correcting itself and writing a working verifier is the kind of reasoning we want to see at this size. Thanks for shipping it. @RunAnywhereAI

RunAnywhere (YC W26) @RunAnywhereAI

2 days ago

2.5B model, offline, solving a puzzle it first called impossible. MiniCPM5 2B from @OpenBMB. 128K context, top open model under 4B when it landed. Searched Hugging Face inside the RunAnywhere app, downloaded, ran.

2 2 28 5K 14

🏥 Bringing local agentic AI to healthcare with MiniCPM5-2B × OpenMed! @OpenMed_AI paired OpenMed with MiniCPM5-2B to explore a local clinical AI workflow — combining privacy-preserving clinical data processing with a compact model capable of tool use, reasoning, and long-context understanding. ✨ Highlights: 🧠 MiniCPM5-2B powers the agent layer, calling tools, comparing lab results, and generating clinical handoffs with source references 🔒 OpenMed masks sensitive identifiers and extracts clinical context before the model processes the data ⚡ Compact 2B-scale model enables practical local inference on resource-constrained hardware 🛠️ Together, they demonstrate how open models can connect clinical data processing with agentic workflows while keeping inference on local hardware It’s exciting to see MiniCPM5-2B move beyond standalone model benchmarks into real-world healthcare workflows — bringing tool use, reasoning, and local deployment together with OpenMed. 🙌 Built something with MiniCPM5-2B? Share your case with us! 🚀 🔗 GitHub: github.com/maziyarpanahi/… 🤗 Model: huggingface.co/openbmb/MiniCP…

keyboard_arrow_left Previous keyboard_arrow_right Next

🎙️ VoxCPM2 meets Turkish voice data! Developer @AlicanKiraz0 fine-tuned VoxCPM2 with nearly 100 hours of natural Turkish speech data, adapting its voice generation capabilities specifically for Turkish text-to-speech. With VoxCPM2 as the foundation, Kahya-TTS explores how open-source voice models can be further adapted to new languages and specialized datasets. ✨ Highlights: 🇹🇷 Turkish TTS powered by VoxCPM2 🎙️ Nearly 100 hours of natural Turkish voice data ⚡ Fine-tuning tailored to Turkish speech synthesis 🛠️ Open-source model adaptation for further research and development Great example of what developers can build on top of VoxCPM2. 🙌 🔗 Model: huggingface.co/AlicanKiraz0/K… 🤗 VoxCPM2: huggingface.co/openbmb/VoxCPM2 Building with VoxCPM2? We’d love to see what you create. 🙌

keyboard_arrow_left Previous keyboard_arrow_right Next

This is exactly the kind of product we hoped VoxCPM2 would enable — speaker-aware cloning, multilingual generation, and a real editing studio around it. Impressive work @dubedostudio 🙌

Dubedo @dubedostudio

4 days ago

Your content. More languages. Your voices. 🎙️ Meet Dubedo: dub into 30 languages, clone each speaker’s voice, and fine-tune the words and timing in one editor. Our beta is open! Try the interactive demo or bring your own clip. 👉 dev.dubedo.com #aidubbing #voicecloning

1 1 1 2K 1

🎙️ What if your videos could speak another language without losing the voices behind them? Developer @dubbedstudio built Dubedo, a commercial AI dubbing studio that helps creators translate and localize existing videos while preserving the voice identity of each speaker. Powered by VoxCPM2, Dubedo generates dubbed speech for each speaker from their translated script and reference voice, bringing the original voices into new languages. ⚡ Natural, expressive speech generation 🎙️ Speaker-specific voice cloning from reference audio 🌍 Multilingual dubbing across 30 target languages 🚀 Cached voice representations + serverless GPU inference From transcription and translation to speaker-aware voice generation and final audio mixing, Dubedo brings the full localization workflow into one place. It’s exciting to see VoxCPM2 powering a real-world product and helping turn open-source speech models into practical applications. 🔗 Try Dubedo: dev.dubedo.com 🤗 VoxCPM2: huggingface.co/openbmb/VoxCPM2 Building with VoxCPM2? We’d love to see what you create. 🙌

We’re grateful to OpenBMB for supporting Atria Dawn Preview with UltraData-SFT-Agent-2609. The dataset’s high-quality agent trajectories have been valuable in developing Atria’s capabilities on agentic tasks. 🔗 huggingface.co/datasets/openb…

@AtriaASI Grateful to see UltraData-SFT-Agent-2609 put to work on Atria Dawn Preview. Congrats to the @AtriaASI team on the preview 👏

🚀 Building natural interaction requires more than a larger model. Gander was trained on ~2.7M interaction examples covering: 🎙️ Speech interaction 👀 Audio-visual interaction 🤖 Agentic interaction 🛡️ Robustness & negative data The training focuses not only on what to say, but also when to listen, when to respond, when to interrupt, and when to delegate. Gander is now open source, with the code, model, technical report, and inference/training setup available for developers and researchers to explore. 🔗 GitHub: github.com/Omni-Interacti… 📄 Technical Report: arxiv.org/abs/2609.08977 🤗 Model: huggingface.co/Gander-Omni/Ga… huggingface.co/openbmb/MiniCP…

⚡ Inside the Cerebellum, Gander builds on MiniCPM-o 4.5’s streaming Thinker–Talker design. The Thinker decides what to do and when to act — listen, speak, interrupt, or delegate. The Talker then generates speech incrementally, allowing perception and conversation to continue while the agent is speaking. Gander also keeps audio, video, and text on a shared timeline, enabling the agent to respond to what you say and what is happening around you.

🤖 What if an AI agent could keep seeing, listening, talking, and working — all at the same time? Developer @speechjsp built Gander, a multimodal duplex interaction agent that combines continuous audio-visual perception, real-time conversation, and asynchronous agent execution. Gander’s Cerebellum is fine-tuned from MiniCPM-o 4.5, with additional training and system-level improvements for native duplex interaction. Instead of treating voice interaction and agent execution as separate steps, Gander brings them together in one continuous interaction loop.

Artificial Analysis gave MiniCPM5-2B an Agentic Index score of 20, and the overall vibe has been very positive for this small model. I tried it locally on llama cpp. At 2B params it’s tiny (the Q8 GGUF is just 2.5 GB), and it handles multi-step tool calls and basic agent tasks without any issue. Gave it two tools and a question needing both: it chained them in the right order, did the intermediate math itself. Great for handing over small background tasks locally. I also tested its multi-step logic on classic reasoning problems (like calculating the thickness of paper folded 42 times). It walked through the full math step-by-step: calculated 2^42 (4.4 trillion multiplier), assumed a 0.1mm sheet, and cleanly converted the result to ~440,000 km (beyond the moon) with zero arithmetic breakdown. It’s blazing fast too (~139 t/s on my M5 Max). 1/3

Insane progress for small language models! MiniCPM5-2B is a dense 2B-parameter model by OpenBMB from China that's 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. I 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 my 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. Download MiniCPM5-2B: huggingface.openbmb.cn/model/openbmb/… A 2B model can now maintain enough state to coordinate tools and complete a useful investigation on local hardware.

@prathamkode 👏Love this use case. A small LoRA on MiniCPM5-2B turning one 2B checkpoint into many game NPCs — stay in character, spoken line only, card in / persona out. That’s the kind of on-device agent we wanted people to build.Nice work @prathamkode

@iamrexei A coding agent in a single browser tab — that’s the edge setup we hoped MiniCPM5-2B would unlock. No Docker, no API key, weights stay local after the first download. Nice work putting Pi + WebGPU + Transformers.

THIS IS no longer just a browser-based chat A coding agent has appeared on Hugging Face that runs entirely within a browser tab, powered by the MiniCPM5-2B model Under the hood, it uses Pi, WebGPU, Transformers.js, and a 4-bit ONNX version of the model Upon the first launch, the browser downloads about 1.8 GB of model weights. After that, the agent runs locally: it can reason, call basic tools, and edit files it has been granted access to The most interesting thing here isn't the model size In the past, a "local agent" almost always implied Docker, Node, Python, a model server, and a complex setup process Here, you simply open a page and get a virtual terminal with an agent ready to go This could be useful for: ▸ quick edits in a small project ▸ experimenting with agents without needing an API key ▸ working with code you don't want to send to a cloud API ▸ prototyping local browser-based tools However, don't mistake this for a replacement for Codex or Claude Code A 2B model can make mistakes on multi-step tasks; WebGPU requires a modern browser and decent GPU memory; and the browser sandbox prevents the agent from accessing the entire disk without permission Still, it sends a strong signal: A coding agent no longer needs to be a separate application or server—it can simply be a single browser tab.

A coding model MiniCPM5 just ran in my browser, No API. No server. WebGPU. just load the model. - 4-bit ONNX, Transformers.js. - 131K context. - It reads files. - Calls tools. workspace, file tools, shell - It reads your files, uses tools, and answers from a local workspace. - huggingface.co/spaces/victor/… Yes Not a Cursor killer, But it’s the first 2B browser agent that feels usable.