MiniCPM-V-4.7: openbmb's 35B-A3B and 1B vision-language models
OpenBMB released MiniCPM-V 4.7 on Oct 6-12, 2026 — a 35B-A3B mixture-of-experts vision-language model plus a 1B companion, Apache 2.0, with FP8/AWQ/GPTQ/GGUF quants and day-0 chip support (Rockchip, Houmo, Intel, AMD, Arm) aimed at on-device deployment.
Why it matters
First open-source MoE vision-language model engineered to run smoothly on-device: 35B total params at ~3B active per token, decoding 13+ tokens/s in ~3 GB of memory on a phone via lookahead expert scheduling.
What you could build with it
A hardware startup building a warehouse inspection robot could run MiniCPM-V-4.7-1B on-device for real-time visual question-answering over camera feeds, with the 35B-A3B variant handling harder cases via API. The result: shelf audits and defect detection without streaming every frame to the cloud.
Does it hold up?
Usable today with real community evidence: QbitAI ran offline chart-reasoning and academic-paper-to-Markdown tasks on a phone, a community vLLM rig on 2x RTX 3090 hits 356 tok/s single-stream with the DSpark draft, and the cookbook plus ROCm/bloomery engines added 4.7 support within days — though vendor benchmarks remain self-reported and the 35B model still needs a multi-GPU host.
Built with MiniCPM-V-4.7
- OpenBMB releases MiniCPM-V 4.7, series' first MoE modelAI Weekly · Alert-style breakdown of the 35B-A3B MoE and 1B release: 13+ tokens/s in ~3 GB on iPhone 17 Pro Max, Apache 2.0, FP8/AWQ/GPTQ/GGUF quants, day-0 Rockchip/Houmo/Intel/AMD/Arm support; flags the vendor performance claims as self-reported with no published benchmark tables.
- ModelBest Open Sources MiniCPM-V 4.7 for Use on the Phone Itselftech360.tv · Third-party writeup covering QbitAI's hands-on tests: the model handled chart reasoning and converted academic papers into Markdown with the phone offline, and scrutinizes the Lookahead Expert Scheduling technique behind the on-device MoE speed claim.
- hf-model-watcher daily vision-model catalogGitHub (arielshakaramiro) · Community GitHub Actions catalog of new Hugging Face vision models; ranked MiniCPM-V-4.7-35B-A3B and -1B among the week's most-liked vision models (227 and 137 likes within days of upload).
- OpenBMB/MiniCPM-Vgithub · 26,545 ★ · Official repo (26.5k stars) with the MiniCPM-V 4.7 model guide, inference, fine-tuning, local deployment and live-demo code.
- tonyd2wild/MiniCPM-V-4.7-35B-A3B-2x-RTX-3090github · 1 ★ · Community day-1 serving work: int4 35B-A3B on 2x RTX 3090 in vLLM with three day-1 fixes, DSpark speculative decoding (356 tok/s single-stream) and full benchmark results.
- rocm/fastflowlmgithub · 1,951 ★ · AMD ROCm on-device inference engine (1.9k stars); commit on 2026-10-08 renamed model keys and NPU engine classes to minicpm-v-4.7 and added 4.7 weights support.
Learn more
First spotted on hf: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72Mistral Large 4 (Le Chonk): 1.05T-param open-weight MoEMistral's new flagship: a 1.05T-parameter mixture-of-experts multimodal model with 49B active params and 1M-token…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71JetBrains Mellum 2.1: 12B MoE coding model trained for agentsJetBrains released Mellum 2.1, a 12B MoE model (2.5B active parameters, 128K context) under Apache 2.0, trained mainly…models · JEV 0.69DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.