OpenBMB · Big Model for Everyone

OpenBMB is an open-source large model community. We build MiniCPM, AgentCPM, VoxCPM and UltraRAG — efficient models that run everywhere, from cloud to device.

Models

Latest

2026-11-03 How far are we from trusting AI with our work?：At the CNCC on-device model forum, Renmin University associate professor Lin Yankai draws on AtomMem, SimpleMemVLA, VeriCUA and related work to examine how agents can learn memory, verification, attribution and correction on their own to complete long-horizon tasks reliably.
2026-10-30 CoDI: Teaching large models to learn at deployment time：CoDI, an EMNLP 2026 SAS Highlights paper from THUNLP and partners, turns meta-induction into a trainable objective for better OOD generalization.
2026-10-28 New SOTA in OCR accuracy and efficiency: ModelBest open-sources MiniCPM-OCR and the first document AI inspector：ModelBest and OpenBMB open-source MiniCPM-OCR 1.0 (1.3B) and Jev, setting new SOTA in efficient document parsing and text recognition.
2026-10-26 Passing the gate doesn’t mean procurement is approved: how MiniCPM5-2B safeguards high-stakes reports：Spark Program: Liu Liang’s ClaimLedger, an evidence-verification Skill for high-stakes reports, runs on Intel AI PCs with MiniCPM5-2B.
2026-10-25 As small models get stronger, is the cloud becoming a strategist? Tsinghua and other teams systematically study Agent device-cloud collaboration：Northeastern University, Tsinghua, and other researchers introduce EC²-Bench, comparing four device-cloud Agent collaboration strategies on 128 tasks.
2026-10-23 Easy to deploy, stable to execute, sustainable to operate: StaffDeck 1.0.0-preview is now open source：StaffDeck 1.0.0-preview is open source, upgrading deployment, execution, and operations so digital employees connect to enterprise context and keep improving.
2026-10-22 Cameras that write daily reports, robots that rap: MiniCPM Spark Program S1 wraps up：Developers turned on-device models into home vision assistants, desktop tower radars, robot performers and dubbing tools in Spark Program S1.
2026-10-21 The next stop for on-device intelligence: a recap of the Personal AI × Physical AI closed-door exchange：OpenBMB and ModelBest gathered five hardware teams in Shenzhen to discuss product form, personal memory, scenario value and the road to Personal AI.
2026-10-21 CNCC 2026: on-device model forum opens tomorrow to explore new frontiers of on-device intelligence：The CNCC 2026 forum gathers Tsinghua and ModelBest researchers to discuss device-cloud collaboration, agents, chip-level customization and speech generation.
2026-10-20 ModelBest Lookahead Expert Scheduling (LES): Running a 35B MoE On-Device in Real Time, Losslessly：How LES runs a 35B-A3B MoE losslessly in real time on a phone at over 20 tokens/s, via scheduling, storage, and compute optimizations.
2026-10-20 OpenSQZ Glass Harness × MiniCPM-o: more controllable, customizable and fluid on-device omni-modal models in the first-person world：OpenSQZ Omni Harness adds interaction control, Session switching, open Skills and a dedicated vision sidecar on top of MiniCPM-o.
2026-10-19 Why Does Multimodal Retrieval ’Favor’ Text? We Found Chaos in the Text：Mixed-modality retrievers systematically favor text over images. We trace the cause and propose Trident so evidence from different modalities competes fairly.
2026-10-17 More Accurate Simulation Does Not Mean More Equal Representation：Are large models better at ’understanding’ Europe and America? A study of 9 open-source models across 59 countries measures representational equality.
2026-10-16 Spark Talk | VoxWeft Pitfall Log: A Development Diary of an On-Device Simultaneous Interpretation Project：Inside VoxWeft, a local real-time interpretation system built on VoxCPM2: a developer’s DEVLOG of deadlocks, crashes, and hard-won fixes.
2026-10-15 The Harness ’little powerhouse’ arrives: PilotDeck 2.0 goes open source with one-click setup and reliable delivery on long, hard tasks：Tasks stall halfway and one small edit forces a full redo. PilotDeck 2.0 turns AI from a chat box into a workbench that truly delivers editable results.
2026-10-13 A new ceiling for on-device multimodal AI: 35B-parameter MoE runs smoothly in 3GB of memory on phones and tablets, MiniCPM-V 4.7 is now open source：MiniCPM-V 4.7 ships in 35B-A3B and 1B versions, with the flagship running losslessly on phones and tablets in about 3GB of memory.
2026-10-12 Tsinghua team revisits OPD: data is stuffed, the algorithm starves：Cutting OPD’s training set from 17,000 problems to one still recovers most of the full-data gain — the bottleneck is how fast the algorithm absorbs supervision.
2026-10-12 Self-Replay Fine-Tuning: Extending Model Capabilities Made Simple：Sample once, reuse across experiments, and extend capabilities with standard SFT while reducing forgetting.
2026-10-09 Event | Tsinghua University Institute for AI annual conference and Beijing-Shanghai-Hong Kong AI exchange opens tomorrow：Tsinghua’s AI institute annual conference and the Beijing-Shanghai-Hong Kong AI academic exchange will be held on campus on October 10-11.
2026-10-09 Closed-door exchange | The next stop for on-device intelligence: Personal AI × Physical AI：What does it take to build a great piece of Personal AI hardware?