MiniCPM-V 4.7 gets a fresh GGUF menu for local vision
mradermacher's new static quantizations add downloadable size options and separate vision projectors for the 35B-A3B model.
Storyline / Report 2 of 2
MiniCPM-V 4.7 local deployment
Full timeline 2 reports
- Community reportsMiniCPM-V 4.7 puts a 35B vision MoE on phones, OpenBMB says
- Community reportsMiniCPM-V 4.7 gets a fresh GGUF menu for local vision (current report)
More download choices for local multimodal users
Community quantizer mradermacher has uploaded a new set of static GGUF builds for MiniCPM-V 4.7 35B-A3B, giving local users more file-size choices for OpenBMB’s vision-language model. The repository was created on October 14 at 05:07 UTC, with successive weight uploads through 06:25 UTC. This is a new community distribution of the model released on October 12, rather than another base-model launch.
The inspected English card lists Q2_K at 13.0 GB, Q3_K_M at 16.9 GB and Q4_K_S at 20.0 GB, alongside larger Q6_K and Q8_0 builds. The quantizer labels Q4_K_S “fast, recommended”; that is the uploader’s recommendation, rather than a comparative benchmark run by AITrending. Separate multimodal projector files are listed at 0.8 GB for Q8_0 and 1.2 GB for f16.
File size is only part of the deployment decision
These are useful options to inspect for people testing local image and video understanding who want a different balance of storage and precision. The listed sizes describe downloadable files, not total runtime RAM or VRAM. Users still need a runtime that supports this architecture and its vision projector; choosing a smaller file does not establish compatibility or quality for a particular device.
The card says weighted or importance-matrix quantizations are not currently available from this uploader. It also provides no new device-speed measurements. OpenBMB’s separately reported phone results should therefore not be assumed to apply to these static GGUF files.