Atlas

Models

← All models

MiniCPM-V 4.6

OpenBMB · 2026-05-11 · 1.3B parameters

MiniCPM-V 4.6 is OpenBMB's compact 1.3B-class vision-language model, open-sourced on May 11, 2026 and built from a SigLIP2-400M vision encoder and Qwen3.5-0.8B language backbone. It uses LLaVA-UHD v4's intra-ViT early-compression technique and mixed 4x/16x visual-token compression, accepts text plus single images, multiple images, and video within a 262,144-token context, and has mobile deployment support for iOS, Android, and HarmonyOS. This repository contains the instruct/non-thinking checkpoint; OpenBMB releases long-chain-of-thought weights separately as MiniCPM-V 4.6 Thinking.

Benchmark scores

BenchmarkScore
AA-LCR6.3
AA-Omniscience Index-84.7
Artificial Analysis Agentic Index1.4
Artificial Analysis Coding Index0.7
Artificial Analysis Intelligence Index4.2
Artificial Analysis Omniscience Accuracy6.2
Artificial Analysis Omniscience Hallucination Rate96.8
Artificial Analysis Openness Index44.4
Capability116.0
CritPt0.0
GDPval-AA v2-83.2
GDPval-AA v2 (normalized)0.0
GPQA Diamond30.5
Humanity's Last Exam (Text-Only)4.9
IFBench (Artificial Analysis)26.7
MMMU Pro37.9
SciCode (Artificial Analysis)2.1
Terminal-Bench 2.10.0
Terminal-Bench Hard0.0
τ²-bench (Telecom)87.7
𝜏³-Banking4.3
Loading Atlas data…