MiniCPM-V 4.7: A 35B AI Model That Runs on Your Phone

Yesterday's other big open release wasn't a giant — it was a shrinking act. On October 15, 2026, ModelBest and the OpenBMB community released MiniCPM-V 4.7, a 35-billion-parameter mixture-of-experts model designed to run on a phone with about 3 GB of memory, decoding at over 13 tokens per second. A 1B sibling covers entry-level devices. It reads images, documents, charts, and long video — offline, on the device, with your documents never leaving your pocket. This is the other end of the cost curve from the datacenter giants: not cheaper per token, but zero tokens at all. Here is what ModelBest shipped, what its benchmarks actually say, the honest catches early testers found, and how to run it yourself.

The short answer

A 35B model that fits in your phone. MiniCPM-V 4.7 comes in two sizes: a 35B-A3B mixture-of-experts model that activates roughly 3 billion parameters per token (the rest waits on SSD storage and loads as needed), and a 1B model for lightweight edge devices. Both are open-source under Apache 2.0 — weights, code, and docs via the OpenBMB project.

The trick is architectural, not magic: sparse activation means the phone only ever runs ~3B parameters at a time, while ModelBest's Lookahead Expert Scheduling (LES) predicts which experts the next layer will need and pre-loads over 80% of them during current computation — with outputs identical to non-scheduled inference, per ModelBest. The result, on OpenBMB's own runtime: 13+ tokens per second on an iPhone 17 Pro Max.

The use case this family is built for: private work on local documents that never leaves the device — chart reasoning, converting papers to Markdown, video understanding, all with the phone offline. That is the pitch, and independent testers confirm the offline part works.

Sources: PRNewswire: ModelBest and OpenBMB Launch MiniCPM-V 4.7 (official announcement, Oct 15, 2026); Tech360: ModelBest Open Sources MiniCPM-V 4.7 for Use on the Phone Itself; Michael Leung: MiniCPM-V 4.7 1B on an 8GB GPU — independent test (Oct 2026). Primary source: the ModelBest/OpenBMB announcement.

What ModelBest actually shipped

Compare the philosophy with Muse Spark Open (released the same day): Meta's 816B model needs a terabyte of memory and a custom license; MiniCPM-V 4.7 needs 3 GB and Apache 2.0. Different universes, same release week — the industry is splitting into datacenter brains and pocket brains simultaneously.

The benchmarks — and what they actually measure

All figures below are OpenBMB's published results (vendor data, transcribed from the official model card). They measure vision-language skill — not coding, not agentic tool use. Keep that frame; it matters for the catches section.

ModelBest's headline claims: in Instruct mode it outscores Qwen3.6 35B-A3B overall while reasoning faster, and the 1B version leads Qwen3.5 0.8B on score and inference efficiency. Strong numbers — for a model you can run on a phone. The honest context: these are the benchmarks OpenBMB chose to publish, and no independent lab has re-run them yet.

The honest catches

• Ollama doesn't support it yet. Ollama lists the model and downloads it, but on 0.40.4 every request fails with "Failed to load CLIP model" — the vision component uses a new format. You need a llama.cpp build from October 10, 2026 or later (independent tester confirmed).
• The 1B isn't a leap over 4.6. Independent testing (RTX 4060 Ti 8GB, ~240 tok/s) found the 1B version no faster and no more accurate than MiniCPM-V 4.6 on the same document jobs. Upgrade if you're new; don't rush if you already run 4.6.
• The 35B needs real hardware on desktop. About 22 GB in Ollama — too much for an 8GB card. The "runs on a phone" claim uses OpenBMB's own mobile stack with experts paged from SSD; desktop GPU setups don't get that trick for free.
• Vision benchmarks ≠ agent benchmarks. Nothing published measures tool use, function calling, or multi-step agentic coding for this model. If your workload is agents that act — not documents that get read — look at the models in our cost-per-task ranking instead. Don't buy a vision model for an agent job.
• Vendor speed figures are best-case. 13+ tok/s is OpenBMB's runtime, their quantization, their speculative sampling, on a flagship iPhone. Your mileage on other hardware will differ.

What to do this week

• Privacy-first document work: this is the model to test. Medical, legal, or financial documents you won't upload to any cloud — MiniCPM-V 4.7 reads them on-device, offline. That's a real, defensible use case today.
• Skip Ollama for now; use a current llama.cpp build (October 10, 2026+) with the model + mmproj files if you want it on desktop.
• Don't upgrade from 4.6 yet unless you need the 35B's MoE efficiency — the 1B's gains over its predecessor look incremental in independent testing.
• Watch for independent benchmarks. When third parties re-run the vision suite — and especially if anyone tests tool use — we'll update this page.
• For agents, not documents: the Muse referral bonus (1B free tokens, code S5V9L9, 48h to redeem) remains the cheapest way to run real agentic workloads — details below.

Frequently asked questions

What is MiniCPM-V 4.7?

MiniCPM-V 4.7 is an open-source (Apache 2.0) multimodal AI model family released October 15, 2026 by ModelBest and the OpenBMB community. It comes in two sizes: a 35B-A3B mixture-of-experts model (~3B active parameters per token) and a 1B model. Both handle text, images, documents, charts, and video, designed to run on-device — including phones.

Can MiniCPM-V 4.7 really run on a phone?

Per ModelBest: yes, the 35B-A3B runs with about 3GB of memory at 13+ tokens/sec on an iPhone 17 Pro Max, using their own runtime with Lookahead Expert Scheduling (most experts stay on SSD and load as needed). The 1B version targets entry-level devices. These are vendor figures on flagship hardware.

Is MiniCPM-V 4.7 free and open source?

Yes — genuinely. Weights, code, and docs are published under Apache 2.0, unlike the custom licenses on some recent 'open' releases. You can use it commercially within Apache 2.0 terms.

Does MiniCPM-V 4.7 work with Ollama?

Not yet (as of mid-October 2026). Ollama downloads it but fails with a CLIP model error; the vision component uses a new format. Use a llama.cpp build from October 10, 2026 or later instead.

Is MiniCPM-V 4.7 good for AI agents?

Unproven. Published benchmarks cover vision-language tasks (documents, OCR, charts, video) — not tool use, function calling, or agentic coding. For agent workloads, choose a model with published agent benchmarks instead.

Should I upgrade from MiniCPM-V 4.6?

Independent testing found the 1B version no faster or more accurate than 4.6 on document tasks. If you already run 4.6, there's no urgent reason to switch; new users should start with 4.7.

Try the Pareto winner free

Muse Spark 1.3 topped our intelligence ranking. Try it with 1 billion free tokens.

Code: S5V9L9

Muse is Meta's AI agent that finishes real errands — email, forms, bookings, research — and keeps working after you close the app. It scored 48 on the Artificial Analysis Intelligence Index, the highest in our 7-model cost-per-task ranking. Join at muse.ai/join, then open Settings > General > Redeem invite code and enter the code above. You get 1 billion free tokens; the person who referred you gets the same.

48-hour deadline: the code must be redeemed within 48 hours of joining, or the tokens are gone for good. Do it the same day.

Claim 1B free tokens with S5V9L9

Full step-by-step: our Muse review.

Want the short version for a specific app? Start with our full directory of 13 reviewed tools and apps or see how we rank them.