#MiniCPM-V
“面壁小钢炮”,全球最强端侧多模态大模型MiniCPM-Llama3-V 2.5

资源url:https://pan.quark.cn/s/b6bb90db464a

海量资源频道:https://t.me/okpojie
海量资源网站:https://www.okpojie.com/
November 13, 2026 at 8:30 PM
“面壁小钢炮”,全球最强端侧多模态大模型MiniCPM-Llama3-V 2.5

资源url:https://pan.quark.cn/s/b6bb90db464a

海量资源频道:https://t.me/okpojie
海量资源网站:https://www.okpojie.com/
November 13, 2026 at 3:00 PM
Can an 8 GB phone from 2020 run a 35B AI model? I tried MiniCPM-V 4.7 35B-A3B from @OpenBMB on my Galaxy S20. It's a mixture of experts (~3B active per token, the rest streamed from storage). After two loader fixes: one word in ~10 s, CPU only. Text only so far. Video: youtu.be/jmQFEj1otho
I Put a 35 Billion parameter LLM on a 6-Year-Old Phone
YouTube video by ML Guy
youtu.be
November 9, 2026 at 9:14 PM
You gotta check this model, it’s 35 billion parameters, pretty large. But it effectively only uses about 3 billion.

Making this model really fast, and efficient. Even running on phones.

ollama.com/openbmb/mini...
openbmb/minicpm-v4.7:35b-a3b
MiniCPM-V 4.7 — the latest multimodal LLM in the MiniCPM-V series, in 1B and 35B-A3B (MoE) variants. Single-image, multi-image, and video understanding, with smooth on-device MoE inference.
ollama.com
November 8, 2026 at 12:15 AM
Thirteen new vision models are now trainable on Nistivo: Qwen3-VL, Qwen2.5-VL 72B, GLM-4.6V and GLM-4.6V-Flash, InternVL3.5, LLaVA-OneVision, PaliGemma 2, MiniCPM-V, and Mistral Small 3.2. The vision shelf grows from 8 to 21, and the catalog holds 108 base models. nistivo.com
November 6, 2026 at 11:52 PM
Posted NVFP4 GGUFs for Gemma-4-31B, MiniCPM-V-4.7-35B-A3B, Nanbeige4.3-33B. Built for 32GB VRAM or less. If it's worth it on consumer cards, I'll make the family. Ping me with suggestions.

https://www.esatapedico.com/posts/latest-model-uploads-gemma-minicpm-nanbeige/

#ai #inference #open-source
November 3, 2026 at 12:10 AM
(not fast but they intend to use it to train smaller)

MiniCPM-V also looking good too. Just lots of incredible medium models.
huggingface.co/openbmb/Mini...
openbmb/MiniCPM-V-4.7-35B-A3B · Hugging Face
Image-Text-to-Text model with 35B parameters. Download, explore files and run it on Hugging Face.
huggingface.co
October 31, 2026 at 2:10 AM
by comparison minicpm-v-4.7-35a3b thinks for a couple seconds and then just starts writing code
October 30, 2026 at 10:22 AM
MiniCPM-V-4.7 ships 65.59 GiB of BF16 weights. Of the cards we launched, the RTX PRO 6000 Blackwell was the one that finished the test, at 1,019 output tokens per second.

https://massedcompute.com/lfm2-5-24b-a2b-minicpm-v-4-7-gpu-buyer-guide/
October 28, 2026 at 3:08 PM
llama.cpp 0.7.0 expands model support and changes server defaults.

Reported gains reach ~2.2× faster decoding and 10× lower embedding VRAM, depending on workload. The default port moves from 8080 to 9931 (Docker stays on 8080 internally); test integrations before upgrading.

Read more on Kimbodo:
What llama.cpp 0.7.0 Changes for Teams Running Open Models in Production
What Happened llama.cpp 0.7.0 expands support for text and multimodal models, including K2 Horizon/MoVA, Prism Bonsai 2, EmbeddingGemma-2, LiquidAI d1, MiniCPM-V 4.7 and new Nemotron models. It also adds speculative-decoding…
kimbodo.com
October 28, 2026 at 9:06 PM
MiniCPM-V-4.7-35B-A3B-4bit, aware of its model string, just told me it was a dense model
i have weird one sided beef with small MoE models apparently
okay but is it any good? best MoE doesn’t really mean much …

why do they always pick such tiny active param counts? are larger ones even worse? i don’t get the tradeoff here
October 26, 2026 at 12:57 AM
llama.cpp now waits when a request targets a busy server slot.

Previously, a prompt-cache update could replace that slot’s context mid-generation—a correctness risk for concurrent requests. The updates also add MiniCPM-V 4.7 multimodal compatibility, not new model weights.

Read more on Kimbodo:
What llama.cpp Updates Mean for Reliable Open-Model Inference
What Happened The documented updates center on llama.cpp rather than a broad wave of new model releases. The most consequential fix prevents a prompt-cache update from replacing the context in…
kimbodo.com
October 23, 2026 at 7:36 PM
MiniCPM-V-4.7-35B-A3B hit 1,019 output tokens per second at 32 concurrent requests on one RTX PRO 6000 Blackwell with vLLM. That is 465 tokens per second per dollar. The numbers are live and public. Launch your GPU on Massed Compute today. github.com/Massed-Compute/gpu-benchmark
October 23, 2026 at 6:01 PM
脱獄PlayStation 5で大型AI(Qwen3.8-27B)を高速動作させるネイティブアプリ「PS5LM Plus」、iPhoneのメモリ3GBで動く35BのAIモデル「MiniCPM-V 4.7」など生成AI技術5つを解説(生成AIウィークリー) www.techno-edge.net/article/2026...
脱獄PlayStation 5で大型AI(Qwen3.8-27B)を高速動作させるネイティブアプリ「PS5LM Plus」、iPhoneのメモリ3GBで動く35BのAIモデル「MiniCPM-V 4.7」など生成AI技術5つを解説(生成AIウィークリー) | テクノエッジ TechnoEdge
「生成AIウィークリー」(第164回)は、Metaが発表した米国製で過去最高性能とうたうオープンウェイトモデル「Muse Spark Open」や、コーディング系評価でGPT-5.6 SolやOpus 5を一部上回る精度のAIモデル「DeepSeek-V4.1-Flash-1017」を取り上げます。
www.techno-edge.net
October 24, 2026 at 12:32 AM
🤗 openbmb/MiniCPM-V-4.7-35B-A3B: 57.9k downloads/30d, Apache-2.0. This 35B vision model is climbing fast. See what's driving its rise.

https://olud.ai/#leaderboard

#OpenSource #AI #HuggingFace
October 19, 2026 at 6:00 PM
⚡ Das Multimodal-Modell für Entwickler.

🛠️ openbmb bringt mit MiniCPM-V-4.7-35B-A3B-gguf ein starkes Modell für die Kombination von Bild und…

🚀 Entdecke diesen und weitere KI-Geheimtipps auf Underground AI.
👉 https://underground-ai.bolt.app/

#KuenstlicheIntelligenz #OpenSource
October 19, 2026 at 2:40 PM
今日のHuggingFaceトレンド

openbmb/MiniCPM-V-4.7-35B-A3B
MiniCPM-V 4.7は、画像や動画の理解が可能なマルチモーダル大規模言語モデルです。
シリーズ初のMoEモデルである35B-A3Bを採用し、効率的な視覚エンコーディングとオンデバイス推論を実現しています。
リソース制限のあるデバイスにおいて、高い性能と推論効率の両立を目指して作成されました。
openbmb/MiniCPM-V-4.7-35B-A3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 19, 2026 at 4:48 PM
MiniCPM-V 4.7 (github.com/OpenBMB/Mini...) brings a 35B-A3B multimodal MoE model to an iPhone at 13+ tokens/s. Its key trick is Lookahead Expert Scheduling: predict which experts the next layer will need and preload their weights while the current layer computes.

#MiniCPMV #MultimodalAI
GitHub - OpenBMB/MiniCPM-V: See, hear, and interact on your local devices.
See, hear, and interact on your local devices. Contribute to OpenBMB/MiniCPM-V development by creating an account on GitHub.
github.com
October 19, 2026 at 4:02 AM
CoreAI Zoo
CoreAI Zoo puts the open-weight AI model zoo on your iPhone, iPad and Mac. Pick a model, download it once, and run it entirely on your device. No account, no server, and nothing you type, say or photograph ever leaves the device. CHAT Stream answers from open language models such as Qwen3, Gemma, LFM2.5, Granite and Mistral, from 0.6B to 30B+ parameters depending on your device's memory. Vision-language models (Qwen3-VL, MiniCPM-V, LFM2.5-VL) answer questions about your photos. VISION Ask questions about an image, read text with on-device OCR, and describe what the camera sees. AUDIO Transcribe speech with Whisper large-v3-turbo, Parakeet and Qwen3-ASR. Talk to an audio-understanding model, generate speech with Kokoro and VoxCPM, and try on-device music generation. CAMERA Live depth estimation, object detection (YOLOX, RF-DETR), action recognition and super-resolution on the camera feed. TOOLS Experimental models such as diffusion-based language models. BENCH Measure load time, tokens per second and memory for any chat model on your own hardware. HOW IT WORKS Every model runs through Core ML on Apple silicon. Weights are downloaded from the Hugging Face Hub on first use over HTTPS, then cached, so the app works offline afterwards. One model is resident at a time to stay within memory limits. New models arrive through a live catalog without an app update. GETTING STARTED Start with Qwen3-0.6B (454 MB) for the fastest first run. Larger models state the free space and memory they need; a recent iPhone Pro or an Apple silicon Mac runs the mid-size ones comfortably. Model weights follow their own licenses, listed on each model's Hugging Face page. The app itself is BSD-3-Clause and part of the open CoreAI model zoo. Report issues and request models at github.com/john-rocky/coreai-model-zoo.
applevis.com
October 18, 2026 at 5:49 PM
今日のHuggingFaceトレンド

openbmb/MiniCPM-V-4.7-35B-A3B
MiniCPM-V 4.7は、単一・複数画像および動画の理解を目的としたマルチモーダルモデルです。
シリーズ初のMoEモデルである35B-A3Bを提供し、LES等の技術により性能を維持したまま、リソース制限のあるデバイス上で効率的な推論を実現することを目的としています。
openbmb/MiniCPM-V-4.7-35B-A3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 17, 2026 at 1:31 PM
MiniCPM-V 4.7 running locally at 100 tokens/second!

New multimodal models from @OpenBMB using only ~3 GB of memory.

Below, I showcase how to run it with @llama_cpp's web UI 🧵 https://t.co/tItzIAD77q
October 17, 2026 at 9:23 PM