OpenBMB Releases MiniCPM-V 4.7 in 35B-A3B and 1B Variants
10/13/2026 — 10/13, 18:23·2 sources (all)·3 reports (all)
Key takeaways
- OpenBMB has released MiniCPM-V 4.7, a new on-device multimodal model family available in two variants: a 35B-A3B MoE version and a 1B versio…
Background & analysis
On October 13, 2026, openbmb released MiniCPM-V 4.7 under its MiniCPM-V series, with two variants surfacing the same day. The first report, posted at 01:16, said the series had gained a 35B-A3B version, the first MoE model in the MiniCPM-V family. It keeps the family's single-image, multi-image, and video understanding capabilities while adding stronger visual encoding and more efficient on-device inference, and its weights and gguf files were published on Hugging Face at openbmb/MiniCPM-V-4.7-35B-A3B-gguf.
Roughly 47 minutes later, at 02:03, a second report said openbmb had also released a 1B version of MiniCPM-V 4.7 aimed at on-device inference. That variant likewise builds on single-image, multi-image, and video understanding, and is described as combining stronger foundation capabilities with efficient visual encoding and on-device inference; its GGUF files were also made available at openbmb/MiniCPM-V-4.7-1B-gguf. The same report mentions a new Pareto frontier in performance and efficiency for this release.
Both posts came from LocalLLaMA and both describe MiniCPM-V 4.7 as the latest model in the MiniCPM-V series, citing the same core capabilities: single-image, multi-image, and video understanding, efficient visual encoding, and on-device inference. The available information ends with the release of these two variants, 35B-A3B and 1B, and the publication of their GGUF weights; the reports give no further updates beyond that point.
AI-generated from 2 reports · updated 17 hours ago
Latest turnModelBest has released MiniCPM-V 4.7, a new family of on-device multimodal models offered in 35B-A3B and 1B versions. The larger model uses an MoE architecture with 35B total parameters but activates only about 3B per inference; the company says it needs roughly 3GB of RAM to decode more than 15 tokens per second on an iPad Pro, and calls it the first open-source MoE model to run losslessly and smoothly on-device.
- Heat index
- 153
- All sources
- 2
- All reports
- 3
- First seen
- 18 hours ago
Reports on this story headlines open the original
ModelBest has released MiniCPM-V 4.7, a new family of on-device multimodal models offered in 35B-A3B and 1B versions. The larger model uses an MoE architecture with 35B total parameters but activates only about 3B per inference; the company says it needs roughly 3GB of RAM to decode more than 15 tokens per second on an iPad Pro, and calls it the first open-source MoE model to run losslessly and smoothly on-device.
华尔街见闻AI score 62openbmb released a 1B variant of MiniCPM-V 4.7, the latest model in its MiniCPM-V series. It keeps single-image, multi-image and video understanding while pairing stronger foundation capabilities with efficient visual encoding and on-device inference, and GGUF weights are already available.
Reddit · LocalLLaMAAI score 52openbmb has released MiniCPM-V 4.7 in a 35B-A3B variant, the first MoE model in the MiniCPM-V series. It keeps the family's single-image, multi-image, and video understanding abilities while adding stronger visual encoding and more efficient on-device inference; weights and gguf builds are available on Hugging Face.
Reddit · LocalLLaMAAI score 62
Other stories people are talking about
- Heat index 1939RisingMeta launches Muse Wonder and Muse Wonder Full models28 sources
- Heat index 625SurgeCharacter.AI Files for Chapter 11 Bankruptcy and Plans Asset Sale to Disney7 sources
- Heat index 607RisingDeepSeek reportedly to lead funding in robotics startup founded by ex-Tencent scientist9 sources
- Heat index 592SurgeAnthropic New Security Measures Cut Model Distillation Data Theft by 90%+7 sources
- Heat index 535RisingEpic CEO confirms major shift to AI-driven game development7 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion
