Intel / article

Openbmb Launches Minicpm V 4 6 1 3b Instruct

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published May 11, 2026

Context

OpenBMB, a China-based lab founded in 2022 by Tsinghua University's NLP Lab and ModelBest Inc., has released MiniCPM-V 4.6 1.3B Instruct, a small vision-language model that accepts text, image and video and replies in text. Its weights are on Hugging Face under an Apache 2.0 license. Artificial Analysis notes that tiny models like this are useful for efficient and on-device use, and that video input at this scale is uncommon (artificialanalysis.ai).

Artificial Analysis scores the non- 13 on its Intelligence Index, the highest for any open-weights model under 2B parameters. That is 3 points ahead of Qwen3.5 0.8B (Non-reasoning, 10) and 2 points behind Qwen3.5 2B (Non-reasoning, 15), which has 1.7 times as many parameters. The closest sub-2B rival, Qwen3.5 0.8B in reasoning mode, scores 11 while using 43 times as many output tokens. MiniCPM-V 4.6 needed just 5.4M output tokens to run the Index, the lowest Artificial Analysis has measured for any open-weights model under 4B parameters scoring 10 or above. On MMMU-Pro, a visual reasoning test, it scores 38%, ahead of LFM2.5-VL-1.6B at 27% and Qwen3.5 0.8B (Non-reasoning) at 26%.

Knowledge is the weak point. It scores -85 on AA-Omniscience, which tests knowledge recall and penalizes , in the same range as other sub-2B non-reasoning models such as Qwen3.5 0.8B (Non-reasoning) at -89 and Exaone 4.0 1.2B (Non-reasoning) at -83. No inference providers were confirmed at release. Two weeks later OpenBMB released MiniCPM5-1B, a text-only model that Artificial Analysis treats as this one's successor.

Key takeaways · AI-distilled

MiniCPM-V 4.6 1.3B is dense, so total and active parameters are both 1.3B. It sits 3 points above Qwen3.5 0.8B (non-reasoning) and 2 below Qwen3.5 2B (non-reasoning), a new Pareto point for intelligence versus parameters.
It scores 38% on MMMU-Pro, the highest visual reasoning score Artificial Analysis has measured for an open-weights model under 2B, ahead of LFM2.5-VL-1.6B at 27% and Qwen3.5 0.8B at 26%.
Its 5.4M output tokens to run the index is about 19x fewer than Qwen3.5 0.8B non-reasoning (101M) and the lowest Artificial Analysis has measured for any open model under 4B scoring 10 or more.
Knowledge recall is weak, as with other sub-2B models: its AA-Omniscience score is -85, near Qwen3.5 0.8B at -89. It has a 262K , BF16 weights and no confirmed API providers at release.

Terms in this piece · Glossary

inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source artificialanalysis.ai