MiniCPM-V-4.7-1B
A 1B on-device vision-language model with single-image, multi-image, and video understanding for phones and edge chips.
About
MiniCPM-V-4.7-1B is the compact variant of OpenBMB's latest MiniCPM-V release, built around a new performance-efficiency pareto frontier for multimodal models. It uses intra-ViT early compression (from LLaVA-UHD v4) to shrink visual tokens inside the vision encoder, with mixed 4x/16x visual token compression so developers can trade visual detail against inference cost. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency in Instruct mode — a vendor comparison worth independent confirmation. It supports iOS, Android, and HarmonyOS plus day-0 chip support for Rockchip, Houmo, Intel, AMD, and Arm, and works with SGLang, vLLM, llama.cpp, and Ollama. Released 2026-10-06 under Apache 2.0, it has about 8.8k downloads and 369 likes.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- vs Qwen3.5 0.8B: higher overall score and inference efficiency in Instruct mode — provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 8/10 (METRICS tier, metrics-v1). Evidence: OpenBMB's MiniCPM line has a multi-year release track record; Apache 2.0; architecture details (intra-ViT early compression, mixed token compression) documented; broad chip and framework support day one. Docked for: (1) headline comparison vs Qwen3.5 0.8B is vendor-run, no independent numbers seen; (2) 1B variant is brand new. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · HF: 369 likes, ~8.8k downloads, created 2026-10-06
- · manual review: OpenBMB MiniCPM line track record; architecture details documented
- · docked: Qwen3.5-0.8B comparison is vendor-run; 1B variant is brand new