Apache-2.0multimodalRuns locally35.2B (HF safetensors BF16)

MiniCPM-V 4.7: Specs, License & How to Run

OpenBMB on-device multimodal VLM; first MoE in the MiniCPM-V series (35B-A3B + 1B).

Still from the official MiniCPM-V 4.7 iPhone 17 Pro Max demo, showing the on-device Welcome to MiniCPM-V screen — Official project image. Source · License: Apache-2.0.

License: Apache-2.0
Parameters: 35.2B (HF safetensors BF16) (active A3B (variant name; not independently counted))
HF likes: 631 (2026-10-15)
First seen: 2026-10-07

TL;DR. MiniCPM-V 4.7 is a multimodal model from OpenBMB (THUNLP and ModelBest, per the MiniCPM-V README). Parameters: 35.2B (HF safetensors BF16) (active A3B (variant name; not independently counted)).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs.

Developer	OpenBMB (THUNLP and ModelBest, per the MiniCPM-V README)
Type	multimodal
Architecture	MiniCPM-V 4.7 35B-A3B is the series’ first mixture-of-experts checkpoint (vendor README). Pipeline tag image-text-to-text. Vendor also describes Lookahead Expert Scheduling (LES) for on-device decode.
Parameters	35.2B (HF safetensors BF16) · active A3B (variant name; not independently counted)
Modalities	image, text, video, document
License	Apache-2.0 — Apache-2.0 on the Hugging Face card (license apache-2.0) and on github.com/OpenBMB/MiniCPM-V LICENSE (checked 2026-10-15). The 35B-A3B card has no LICENSE blob (404); we follow the card tag plus the code repo LICENSE. Not legal advice.
Hardware	3 GB (vendor, iPhone 17 Pro Max) — The MiniCPM-V README (checked 2026-10-15) claims 13+ tokens/s with 3 GB of memory on iPhone 17 Pro Max for 35B-A3B via Lookahead Expert Scheduling. That is a vendor/self-reported device figure. forkalt has not measured decode speed or peak memory. (not independently tested)
Hugging Face likes	631 (as of 2026-10-15)
Model card	https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B
First seen	2026-10-07

MiniCPM-V 4.7 is an open-weight multimodal model family from OpenBMB. Official hosts we opened on 2026-10-15: the ungated card openbmb/MiniCPM-V-4.7-35B-A3B, the 1B card openbmb/MiniCPM-V-4.7-1B, the collection openbmb/minicpm-v-47, the code repository OpenBMB/MiniCPM-V, and the demo space openbmb/MiniCPM-V-4.7-Demo.

This identification page hangs off ChatGPT alternatives and closed LLM APIs. Those mappings are editorial. We have no sourced quality comparison against ChatGPT.

The README (changelog 2026-10-12) calls 4.7 the latest MiniCPM-V release, available as 35B-A3B and 1B, and says the 35B-A3B variant is the series’ first MoE. Hugging Face pipeline_tag is image-text-to-text. Those are copied labels, not a ranking.

What we verified (2026-10-15)

Field	Value	Source
35B-A3B HF likes / downloads	631 / 4,157	HF API `openbmb/MiniCPM-V-4.7-35B-A3B`
Card created / last modified	2026-10-07T07:07:30Z / 2026-10-14T15:16:30Z	same
Gated	false	HF API
License (card)	apache-2.0	HF `cardData.license`
LICENSE blob on the 35B card	404	HF `.../raw/main/LICENSE`
Code-repo LICENSE	Apache License 2.0	OpenBMB/MiniCPM-V LICENSE
Safetensors params (35B-A3B)	35,212,875,824 BF16 (~35.2B)	HF `safetensors.parameters.BF16`
1B card likes / downloads	333 / 2,543	HF API `openbmb/MiniCPM-V-4.7-1B`
1B created	2026-10-06T17:40:13Z	same
1B safetensors	1,554,476,272 BF16 (~1.55B)	HF API
GitHub stars / forks	26,754 / 2,096	GitHub API `OpenBMB/MiniCPM-V`
README release line	2026-10-12 · MiniCPM-V 4.7 released	GitHub README

Quantized siblings on the same collection (copied ids, not separately re-measured): GGUF, FP8, GPTQ, AWQ, int4, and DSpark heads.

On-device claims (vendor / self-reported)

The MiniCPM-V README says 35B-A3B, with Lookahead Expert Scheduling, is “the first open-source MoE model to run smoothly on-device without compromising performance,” and quotes 13+ tokens/s with 3 GB of memory on iPhone 17 Pro Max. It also claims Instruct-mode 35B-A3B “achieves a higher overall score in less inference time than Qwen3.6 35B-A3B.” Those sentences are vendor/self-reported. forkalt has not re-measured speed, memory, or those benches, and we do not turn them into a ChatGPT verdict.

The page screenshot is a still from the official iPhone demo GIF on the 35B-A3B card (v47_35ba3b_iphone2_en.gif).

How to run (from the official card, not yet tested)

The Hugging Face README (checked 2026-10-15) documents Transformers:

pip install "transformers[torch]>=5.18.0" torchvision torchcodec

and AutoModelForImageTextToText / AutoProcessor from openbmb/MiniCPM-V-4.7-35B-A3B. A live demo is openbmb/MiniCPM-V-4.7-Demo. forkalt has not run these commands.

License

The 35B-A3B card is tagged license: apache-2.0. There is no LICENSE file on that Hugging Face repo (404 on 2026-10-15). The code repository OpenBMB/MiniCPM-V ships an Apache License 2.0 file, and the README states code and model weights are Apache-2.0. We record both the card tag and the missing blob. Not legal advice.

What we did not verify

Vendor Pareto-frontier charts, Qwen comparison tables, and the iPhone 17 Pro Max 13+ tok/s / 3 GB figure stay unlabeled measurements. We have no sourced bake-off against ChatGPT or against DeepSeek-V4.1-Flash.

Related chat weights on this site: DeepSeek-V4.1-Flash on ChatGPT alternatives; Kolibri-1 and LFM2.5-24B-A2B on closed LLM APIs. Same org, different job: MiniCPM-OCR 1.0 (document OCR, not this chat VLM) — MiniCPM-OCR 1.0 vs Unlimited-OCR-2.

Deploy options

local: Hugging Face weights (35B-A3B)
local: Hugging Face weights (1B)
hosted: Hugging Face Space MiniCPM-V-4.7-Demo

Comparisons

MiniCPM-V 4.7 vs Iris-3B

FAQ

Are MiniCPM-V 4.7 weights public?

Yes. Hugging Face openbmb/MiniCPM-V-4.7-35B-A3B is ungated (created 2026-10-07; 631 likes and 4,157 downloads in the 2026-10-15 API check; pipeline_tag image-text-to-text; license apache-2.0). A 1B variant is openbmb/MiniCPM-V-4.7-1B (created 2026-10-06; 333 likes; ~1.55B BF16 params). Code is github.com/OpenBMB/MiniCPM-V. The README changelog dates the 4.7 release 2026-10-12.

How big is MiniCPM-V 4.7?

Hugging Face safetensors on the 35B-A3B card report 35,212,875,824 BF16 parameters (~35.2B). The card name is 35B-A3B; we do not independently count active experts. The 1B card reports 1,554,476,272 BF16 parameters (~1.55B).

Can I run it on an iPhone?

The official README claims 35B-A3B decodes at 13+ tokens/s with 3 GB of memory on iPhone 17 Pro Max using Lookahead Expert Scheduling, and shows device screen recordings. That is a vendor claim. forkalt has not reproduced it. A Hugging Face demo space is also linked.

Does it match ChatGPT quality?

We do not claim that. This page identifies an open-weight on-device VLM. README tables versus Qwen3.6 35B-A3B and Qwen3.5 0.8B are vendor self-reported. There is no sourced bake-off against ChatGPT on this page.

Sources & verification

Last verified 2026-10-15 · Last updated 2026-10-15· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.