Parameters never equaled capability. MiniCPM-V 4.6 from ModelBest (OpenBMB) drives that point home again with just 1.3B parameters.
On May 11, 2026, ModelBest and Tsinghua University released this on-device multimodal model. It weighs only 1.3B total — a 0.8B language model plus the SigLIP2-400M vision encoder. Yet it scores 13 on the Artificial Analysis Intelligence Index, higher than its own language backbone Qwen3.5-0.8B (10), and even above the 3B-class Ministral 3.
This is not "small models that can run" — it's "small models that win."
1. Architecture: Why 1.3B Outperforms 3B
| Component | Model | Params | Role | |------|------|--------|------| | Vision encoder | SigLIP2-400M | 400M | Image feature extraction | | Language model | Qwen3.5-0.8B | 0.8B | Text reasoning, output | | Merger | Window-Attention Merger | - | Visual token compression | | Total | - | 1.3B | Runs on-device |
Two design choices matter:
Compression inside the ViT. Conventional vision encoders compress tokens after the ViT finishes processing. MiniCPM-V 4.6 moves compression into the ViT itself, reducing token count during processing. Compute drops by more than 50%.
Hybrid 4x/16x visual token compression. Switch to 4x mode for high precision, 16x mode for fastest inference. Modes are switchable at runtime — like a gear shifter for the model.
Efficiency numbers (key)
| Metric | MiniCPM-V 4.6 | Qwen3.5-0.8B | Comparison | |------|---------------|--------------|------| | Token throughput | 1.5x baseline | 1x | 50% faster | | Compute token consumption | ~5.4M | ~100M | Only 1/19 | | Memory footprint (Q4) | ~1.6 GB | - | Tiny | | Runtime RAM requirement | 6 GB | - | Phone-capable | | First-token latency | Milliseconds | - | Real-time interaction |
Same hardware, 19x the traffic. That's not optimization — it's restructuring the cost curve.
2. Capabilities
- Single-image understanding: captioning, visual QA, OCR
- Multi-image understanding: cross-image comparison, sequential reasoning
- Video understanding: temporal analysis, frame-level reasoning (128K context supports long video)
- Tool calling: native function calling, usable as an Agent vision module
- Document parsing: invoices, PDFs, screenshots, handwriting
- Thinking mode: optional chain-of-thought reasoning variant
- In-ViT compression (architecture level)
- Hybrid compression ratios (runtime level)
- SigLIP2 vision encoding (pretraining level)
- Qwen3.5 language backbone (language-model level)
- Do complex spatial reasoning ("which object is third from the left")
- Parse dense scientific charts
- Deeply understand long video beyond the context window
- Handle safety-critical scenarios without dedicated safety training
- Offline on-device OCR and document scanning
- Vision perception for on-device agents (reading screenshots, UIs)
- High-concurrency document pipelines (invoices, tables, PDFs)
- Privacy-sensitive scenarios (local medical or financial document analysis)
- Multimodal feature extraction for recommender systems (validated at Kuaishou)
- Team: ModelBest (OpenBMB) + Tsinghua University
- Release date: 2026-05-11
- License: Apache 2.0
- Hugging Face: https://huggingface.co/openbmb/MiniCPM-V-4.6
- GitHub: https://github.com/OpenBMB/MiniCPM-V
- Mobile demos: https://github.com/OpenBMB/MiniCPM-V-Apps
- Web demo: https://huggingface.co/spaces/openbmb/MiniCPM-V-4.6-Demo
- Supported frameworks: Ollama, vLLM, SGLang, llama.cpp, Transformers
- Fine-tuning: LLaMA-Factory, SWIFT
Industrial validation: In 2025, Kuaishou's OneRec recommendation model handled 25% of requests in the short-video main scenario. Not lab data — real load at hundreds of millions of daily active users. MiniCPM-V 4.6 brings that industrially validated compression stack down from 8B to 1.3B.
3. Deployment: From Cloud to Fingertips
Native mobile support
OpenBMB open-sourced all on-device adaptation code, with complete project source for three platforms:
| Platform | Dev environment | Reference device | |------|---------|---------| | iOS | Xcode / Swift | iPhone 17 Pro Max | | Android | Gradle / Kotlin | Redmi K70 | | HarmonyOS | DevEco / ArkTS | HUAWEI nova 14 |
Built on the llama.cpp engine, all inference runs locally: offline, private, millisecond responses. Q4_K_M quantized model files total about 1.6GB (0.5GB LLM + 1.1GB vision module).
Server deployment options
| Framework | Use case | |------|---------| | Ollama | Quick local testing | | vLLM | High-concurrency serving | | SGLang | High-performance inference | | llama.cpp | CPU / embedded devices | | Hugging Face Transformers | Research / fine-tuning |
Quantization formats cover GGUF, BNB, AWQ, GPTQ; FP16 is ~3-4GB, INT4 ~1.5-2GB.
4. Competitive Landscape
| Model | Params | Strengths | Weaknesses | Best for | |------|--------|------|------|------| | MiniCPM-V 4.6 | 1.3B | Extreme efficiency, strong OCR, low cost | Weak complex spatial reasoning | On-device agents, document processing, OCR pipelines | | Qwen3.5-VL-2B | 2B | Alibaba ecosystem, Chinese optimization | Less efficient than MiniCPM | Cloud Chinese-language apps | | Moondream | 1.8B | Minimal deployment | Weak OCR/document | Basic captioning | | LLaVA-Phi | 3.8B | General VQA | Large, inefficient | General scene understanding | | GPT-4o | >100B | Everything | Cloud, costly, privacy risk | High-precision general tasks |
The competitive positioning is clear: don't fight GPT-4o on breadth — beat same-class small models on intelligence density per parameter.
5. Why Small Models Keep Getting Stronger
ModelBest has iterated the MiniCPM-V series from April 2024 to version 4.6 in two years. The core thesis: intelligence density > parameter scale.
Traditional thinking: bigger is better, scale infinitely. On-device thinking: smaller is cheaper, good enough suffices.
MiniCPM-V 4.6 takes a third path: maximize intelligence density within a tiny parameter budget via architectural innovation and training efficiency.
Efficiency is squeezed out at every layer. Not one trick — systematic frugality.
6. Limitations and Fit
Don't expect it to:
It fits best for:
7. Bottom Line
MiniCPM-V 4.6 is not a "cheap substitute for large models." It's an on-device multimodal model that recalibrates the efficiency baseline. 1.3B parameters, 1.6GB model, 6GB RAM, 19x token-cost advantage — these numbers are not compromises; they're a new coordinate system.
If you're building on-device AI, local agents, OCR pipelines, or just want to cut the cost of multimodal serving, this model deserves first-priority testing.