Bonsai 2 4B vs MiniCPM-V 4.7 1B on iPhone. Same photo, 3 building ideas. Which answer would you use? @PrismML @OpenBMB My app: apps.apple.com/app/id6804004… Prompts differ slightly. Waits cut; not a speed test.

🚨 @OpenBMB just dropped MiniCPM-V 4.7, and the engineering behind it is seriously impressive. 35B total parameters, ~3B active at inference, and ~3GB memory in its target on-device configuration. Here's what makes it work: → LES predicts which experts the next Transformer layer needs and preloads their weights to reduce memory latency. Misses get rescheduled to maintain parallelism. → DSpark Speculative Decoding drafts tokens ahead of the main model, which verifies them. Same output, faster execution. → Visual Token Compression cuts image and video token overhead by 4x and 16x. That last one makes a real difference. So I built something to test it 👇 🎬 Clip Narrator: drop in ANY video clip and get back a fully narrated version. My test footage? Literally 10 seconds of milk heating in a pot. The 35B-A3B sampled 8 frames, compressed the clip into ~2,500 visual tokens, and generated a timestamped movie-trailer script: "A pot of milk heats up." "Steam rises. The stakes are high." "It boils over! Disaster strikes!" A TTS model brought it to life, generating 8 seconds of audio in just 2.6 seconds. Subtitles synced automatically. Export and done. Here's what impressed me: it wasn't just captioning 8 still frames. It tracked the milk rising, steam building, and boil-over, timing each line to the action. That's temporal reasoning, not just image description! A 1.3B variant also handles OCR, charts, math, images and video. And both support 256K context, with fully open-source weights, code and fine-tuning pipelines. Beyond scaling parameters, it's about smarter execution on consumer hardware. Try it locally and tell me the most boring clip you'd feed it first 👇 FREE and open-source. Repo link in 🧵↓

我用 ESP32 做飞机姿态仪，还能跟塔台连麦对话 一块 ESP32 开发板 + 触摸屏 + 六轴 IMU，我做了一个拿在手里随便转、画面都跟得上的飞机姿态仪，还能按住按钮跟"塔台"连麦对话。 姿态仪本体：地平线和俯仰梯居中，顶部横滚刻度弧，左空速带、右高度带、底部罗盘，配四个荧光绿读数窗；提示音覆盖拉起、失速抖杆、坡度、超速和整套进近高度喊话。 语音链路五步，全部本地跑： ① 按键说话，ESP32 走 WiFi 把语音传到电脑 ② Qwen3-ASR-1.7B 本地识别，半秒内出字 ③ MiniCPM5-2B 写塔台回复——它会先读姿态仪的实时数据，你人在天上还喊"请求起飞"，直接怼回来 ④ VoxCPM2 音色迁移，拿一段塔台录音做参考，合成同款音色 ⑤ 加电台信道音效，回传设备播放 从松手到塔台开口，大概 2~3 秒。 代码大部分是我和 Kimi Code 对话聊出来的。 项目已开源 ，esp32-attitude仓库：github.com/emwstudio/esp32-a… #esp32 #AI硬件 #创客 #ai语音对话 #飞机姿态仪

MiniCPM-V 4.7 35B-A3B now runs in a browser tab, in LocalMind: it reads images, uses tools, and nothing leaves your device. Big photos go in as an overview + up to 9 tiles, the way OpenBMB trained it. 19.7 GB, streamed from disk. localmind.naklitechie.com

Your Mac just built a game without asking the internet. One prompt. A plan. Three files. Playable Tetris with a ghost piece and hold queue. Pure Swift. No Python. Nothing left the machine. MiniCPM-V 4.7 35B-A3B, fully offline in Osaurus.

MiniCPM-V 4.7 1B, now running fully offline on iPhone in OnDevice LLM. Snapped a coffee receipt, asked "how much was the tip?" and got $0.85. "What's the total?" $23.18. No internet, no cloud. Congrats @OpenBMB 👏

Small models are awesome! The newest OpenBMB model MiniCPM-V 4.7 1B read 11px contract text, rebuilt a table cell-perfect and nailed a blurry, tilted label. On reading, it keeps up with its 35B sibling. Counting still trips it up, but wow. More tests are coming

16 Parallel Agents At The SAME TIME ! 🚀 Minicpm-v-4.7 on MiniCPM-V-4.7-35B-A3B int4 + DSpark draft — 16 parallel coding agents • Total: 59,459 tokens in 29s • Aggregate: 2192.9 tok/s sustained · 2067.7 wall-avg · 2605.4 peak burst • Per-stream decode: 142.1 high / 133.7 low / 137.1 avg tok/s • TTFT 0.27s avg · E2E 27s avg

Mini CPM-V 4.7 brief test! Numbers from the 5090 rig: Q5_K_M, 214 tok/s, 11.2K tok/s prompt processing, 262K context in 30.1 GB, 0.4 s vision. It produced a very decent 3d scene in Claude Code given the speed and model size. It was also able to fix its own errors autonomously! Most interestingly though, it can be run much faster offloaded than prior 35B MOE models. Looking forward to trying it on my MacBook and perhaps even my iPhone! If you're wanting a new Qwen 35B, this might be a fun one to try in the interim!

💥 MiniCPM-V 4.7 35B-A3B just dropped, and we put it through five real tests on an H100 📱 Open-source MoE built for on-device use, now running on a local GPU 🔹 Setup and serving on a single H100 🔹 Coding: a chimney cake fire simulation 🔹 SQL: find the hidden bug and fix it 🔹 Vision: a handwritten note in three languages 🔹 Physics: reasoning over a page of equations 🔹 Translation: one Brazilian headline into 80 languages 🔥 Full test and honest verdict in the video below 👇 piped.video/zFJiwyhDFrA?si=3DJY… @OpenBMB

MiniCPM-V4.7 is out. i got api early access from @OpenBMB and ran both sizes through 13 benchmarks with BenchKit. the 35B-A3B is a really strong vision model. 86 on MMVP, 91 on OCRBench, 90.4 on ChartQA. it's also the best three.js scene writer i've tested so far (88 vs 80 for Empero’s qwen3.8-35b-a3b). the 1B gets 82 on OCRBench and 85 on ChartQA. that's wild for 1.3B. it can't write full programs though. coding and agentic tasks are its weak side. qwen3.6-35b-a3b is still ahead on HumanEval+ and PatchEval. all scores first attempt, official API. weights config coming later this day.

🚀 Tiny model, big test: MiniCPM-V 4.7 1B, hands-on 📱 A one billion parameter vision model that fits on a phone 🔹 Installed the GGUF with llama.cpp and Ollama 🔹 Read an analog clock from a photo 🔹 Turned that clock photo into working HTML code 🔹 Solved a quadratic from a school math sheet 🔹 Handwriting translation across dozens of languages 🔹 Funny scene test: spotting body language and backstory 🔥 Honest results, wins and fails, in the full video below 👇 piped.video/40FXBcRz4_k?si=UGnv… @OpenBMB

🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multimodal capabilities to iOS and Android devices with extreme efficiency. Two new models: 🔹 MiniCPM-V 4.7 35B-A3B — Bring SOTA performance to local devices with newly introduced LES (Lookahead Expert Scheduling) technique, a new milestone of on-device multimodal AI. 🔹 MiniCPM-V 4.7 1B — Lightweight, efficient multimodal AI. Easier to deploy. ⚡ What's new? SOTA performance: Competitive results across 34 benchmarks spanning multimodal understanding, STEM reasoning, OCR and long-video comprehension. Fast on-device inference: LES predicts and preloads expert weights to reduce inference stalls, boosting 5× speedup on iPad Pro M4 (16GB) with byte-for-byte identical outputs. It enables MiniCPM-V 4.7 35B-A3B to achieve 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, while costing only 3GB memory 🤯 Efficient visual processing: 4× and 16× hybrid visual token compression. Even at 16× compression, 35B-A3B scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B-A3B. Easier customization: We introduce Self-Replay Fine-Tuning, a simple way to customize new target capabilities at low cost while maintaining general performance. Broad hardware support: Day-0 adaptations for Apple, Intel, AMD, Arm and Rockchip, with many more platforms supported through FlagOS! Download, deploy, and build something we haven't imagined yet. 👀👇 🔗 GitHub: github.com/OpenBMB/MiniCPM-V 🤗 Hugging Face: huggingface.co/collections/o…

🎬 Excited to share our NeurIPS 2026 research: OmniCapBench – a comprehensive benchmark for evaluating omnimodal video captioning models! 🌐 Homepage: 01yzzyu.github.io/OmniCapBen… 📄 Paper: arxiv.org/abs/2610.12458 🤗 HF Papers: huggingface.co/papers/2610.1… ⭐ Data: huggingface.co/datasets/Omni…

🚀 Meet MiniCPM5-2B, a 2B-parameter language model bringing high intelligence density to the edge, now open source! It ranks #1 among open-source models under 4B parameters on the @ArtificialAnlys Intelligence Index, with a score of 23. It also scores 20 on the Agentic Index, bringing an early form of general-purpose agent capability to the edge. Across 34 benchmarks, MiniCPM5-2B achieves an average score of 53.9, covering coding, math, long-context understanding, tool use, and agentic tasks. And this release goes beyond the model itself. We’re opening up the data, training recipes, and RL stack behind MiniCPM5-2B. 🤗 Hugging Face: huggingface.co/openbmb/MiniC… 💻 GitHub: github.com/OpenBMB/MiniCPM Modelscope: modelscope.cn/models/OpenBMB… Web: openbmb.cn/