OpenBMB

"Scaling Small LLMs to 128K Reasoning with a Critic." The sequel to JustRL (December 2025, arXiv 2512.16649), which argued that a minimal single-stage…

"A Role-Driven RL Training Framework under the SPMD Paradigm." OpenBMB's post-training framework makes every RL role (inference, training, rollout) an…

Second model in the MiniCPM5 series, released September 7, 2026, scaling the MiniCPM5-1B recipe to a dense 2.52B-parameter (1.98B non-embedding) stand…

OpenBMB's September 2026 UltraData drop, released alongside MiniCPM5-2B, whose pretraining corpus it feeds. The public drop is the L2 tier, split by l…

OpenBMB's first embodied series (Apache 2.0): RobotManip, a 1.5B generalist VLA using MiniCPM-V visual-token compression (60-frame streaming at 120ms/…

"Refining Pre-Training Data at Scale with Adaptive Programmatic Editing": instead of filtering or rephrasing, a lightweight refinement model predicts…

The first formal benchmark for theorem proving in mathematical analysis: 200 Lean 4 + Mathlib problems in two tiers (100 undergraduate, 100 PhD-qualif…

Tri-modal foundation model unifying histology imaging, spatial transcriptomics, and biological language for spatial biology and pathology-related reas…

L3 refinement of Ultra-FineWeb — the same L2-curated text rewritten into more learnable training samples with explicit reasoning signals. 400B+…

First release of the MiniCPM5 series — jumps past MiniCPM4 with a 1.08B dense Llama-architecture model featuring 128K context and hybrid thinkin…

First publicly reported 1.58-bit ternary quantization-aware training at 8B scale on Huawei Ascend NPU (CANN stack). Four sizes released: 0.5B, 1B, 3B,…

1.3B-parameter on-device vision-language model with 262K context window and 4x/16x mixed visual-token compression for iOS, Android, and HarmonyOS depl…

Landmark release using Sparse and Linear Attention for 1M-token context windows on a single consumer-grade smartphone.

Most advanced real-time omni-modal agent from OpenBMB. Matches Gemini 2.5 Flash in vision, speech, and live-streaming interaction latency and intellig…

Adds InfLLM-v2 trainable sparse attention to MiniCPM4, supporting both deep reasoning and non-reasoning modes. 3x decoding speedup for reasoning tasks…

State-of-the-art on-device vision model specializing in complex document parsing and high-resolution spatial reasoning.

Ultra-efficient 8B LLM for end devices. Introduced Hybrid Reasoning (dual-mode switching between deep reasoning and fast response) and InfLLM-v2 spars…

OpenBMB's L2-selected open pretraining dataset and the filtering pipeline behind it. ~1.3T tokens — 1T English and 120B Chinese — distille…

A series of high-performance, "edge-scale" agents (4B-8B) designed for long-horizon reasoning, deep research, and mobile GUI automation. AgentCPM demo…

Low-code framework for building complex and innovative RAG pipelines.

Omni-modal model (8B) capable of real-time speech-to-speech interaction and multimodal live streaming on mobile devices. Built on SigLip-400M + Whispe…

Introduces the concept of **Capacity Density**, a metric to evaluate LLM quality across different scales. The paper identifies an empirical "Densing L…

4B dense model optimized for reasoning and RAG, matching performance of 7B-9B models like Llama 3.1-8B while being significantly smaller and more effi…

A breakthrough 8B-parameter multimodal model (built on Qwen2-7B) that surpassed GPT-4V in single-image, multi-image, and video understanding tasks. It…

"RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness." Aligns multimodal LLMs through fully open-source AI feedback, combining high…

Vision-language model series achieving GPT-4V level performance on mobile devices. The first multimodal model deployed natively on a smartphone. Progr…

An efficient end-side Multimodal LLM (MLLM) combining MiniCPM-2.4B with a SigLip-400M vision encoder. It introduced high-resolution image support (up…

The original MiniCPM series proving that small models can rival much larger ones. This work introduced the **Warmup-Stable-Decay (WSD)** learning rate…

Framework enabling LLMs to use over 16,000 real-world APIs.

Multi-agent framework where agents communicate to develop software. ChatDev 2.0 (DevAll) introduces a zero-code multi-agent orchestration platform for…

A 10B parameter foundation model and the second milestone of CPM-Live. It features a unique JSON-structured input/output format for natural multi-task…

High-quality alignment datasets for instruction-following and RLHF.

Open-source framework for building tool-using agents (similar to OpenAI Plugins).

The first milestone of the **CPM-Live** project. A 10B parameter Chinese-English bilingual model trained through a transparent "live broadcast" proces…

Foundational work on parameter-efficient fine-tuning (PEFT), published in Nature 2023. Allows massive models to be updated with minimal data/compute.

Efficient training library for large-scale transformers.