1/5 MiniCPM-V 4.6 (1.3B) is now live 🚀🚀 High-res visual processing, optimized for consumer-grade and mobile hardware. We’ve leveraged the latest LLaVA-UHD v4 technique to cut vision encoding costs by 55%, enabling native edge deployment with extreme efficiency. 🔥 Beats Gemma4-E2B-it and Qwen3.5-0.8B across key multimodal and Artificial Analysis... benchmarks — scoring higher than Qwen3.5-0.8B using just 2.5% of its token budget. ⚡ TTFT (75.7ms) 2.2x Faster than Qwen3.5-0.8B even with 3136² high-res images. 🏗️ ~1.5x Token Throughput compared with Qwen3.5-0.8B on a single RTX 4090. Try the model here: 🤗 Hugging Face: 💻 GitHub: 🔭 Modelscope: 🌐 Web Demo: 📱 App Demo:show more

2/5 The Technique behind the Efficiency: LLaVA-UHD v4 MiniCPM-V 4.6 debuts the new LLaVA-UHD v4 architecture, slashing vision encoding FLOPs by 55.8% without performance degradation. 🔹 Intra-ViT Early Compression: Dramatically reduces visual token counts in early encoding layers. 🔹 Hybrid 4x/16x Visual Compression: Toggles between high-precision (4x compression) and extreme-speed (16x compression) in a single model.

3/5 Extreme Computation Efficiency 📈 Thanks to the LLaVA-UHD v4 technique, MiniCPM-V 4.6 nearly flattens the "Resolution-Latency" curve. Even with 3136² high-res images, TTFT is just 75.7ms on a 4090 GPU — 2.2x faster than Qwen3.5-0.8B. On a single RTX 4090, MiniCPM-V 4.6 achieves ~1.5x token throughput compared with Qwen3.5-0.8B, delivering maximum ROI for high-frequency video/image analysis

4/5 High Token Efficiency MiniCPM-V 4.6 uses 19x fewer tokens than Qwen3.5-0.8B, and 43x fewer tokens than Qwen3.5-0.8B-Thinking to score higher on the Artificial Analysis benchmark. Less compute, broader device coverage, faster inference. That's what efficiency actually looks like. 🚀

5/5 Built for developers 🔹 Deployable on iOS, Android, and HarmonyOS — all edge adaptation code open-sourced, reproducible in just a few steps 🔹 Fine-tuning support on consumer-grade GPUs via SWIFT and LLaMA-Factory 🔹 Compatible with SGLang, vLLM, llama.cpp, and Ollama out of the box 🔹 Multiple quantized formats available: GGUF, BNB, AWQ, GPTQ #AI #EdgeAI #OnDeviceAI #VisionLanguage #OpenSourceAI #MiniCPM #OpenBMB

TTFT is honestly the metric I care about most in production multimodal systems.

The 19x lower token usage compared to Qwen3.5-0.8B is honestly wild.

Cool！

Congrats!👏

Thanks！

"Edge efficiency isn't a consolation prize — it's the next frontier."

Smaller and practical is where AI is heading.

Massive efficiency gains like this are what make multimodal AI actually scalable in the real world.

Efficiency is becoming the real competitive edge.

7ms TTFT on 3136² images with a 1.3B model is honestly wild

This feels much closer to real-world deployment

Smaller models winning on efficiency is going to reshape multimodal deployment at scale.

This is actually impressive for a 1.3B model. Edge AI is moving faster than expected.

The time to first token speed is a massive leap.

Native Ollama support out of the box is exactly how you move from prototype to production.

yeah

thanks congrat

its cool

The efficiency is real, but keep prompts strict for multi-step reasoning at this size

This is cool!

happy to meet MiniCPM-V 4.6 AI model

@grok how do i run this locally on my iphone 17 pro

Smaller, faster, cheaper is the kind of progress that actually changes how people build.

Amazing work by the team. The efficiency gains here are seriously impressive.

Excellent Work 👏🏽

The OCR capabilities on this version look incredibly solid.

sure! give it a try!

Efficiency + Performance = The future of on-device AI. Great work team!

thanks for support!

The local AI ecosystem is getting really interesting.

Native Ollama support makes integration so much smoother.

yeah!

This is a game changer for low end devices.

Great work OpenBMB. This is huge for on-device AI.

Congratulations 🎈

thank you 🤩

The single GPU fine tuning is a huge relief.

1.3B modèles sont rapides, mais la latence dépasse 300ms au-delà de 4k tokens.

This level of quality always gets attention 👀

This project has huge potential 🚀 would love to collaborate and contribute