MiniCPM-V 4.7: MoE model beats Qwen3.6 on doc understanding, 1B runs in 2.9GB RAM
NielsRogge · x · 2026-10-17
OpenBMB released MiniCPM-V 4.7 in 35B-A3B (first MoE in the series, 4x compression) and 1B variants for on-device multimodal use.
- Official benchmarks vs Qwen3.6-35B-A3B: OmniDocBench v1.6 94.2 vs 80.2, CharXiv reasoning 75.6 vs 67.5, OCRBench 91.2 vs 86.0 — notably strong at documents and charts.
- The 1B model needs just 2.9GB RAM and runs fast, per Niels Rogge, with Lookahead Expert Scheduling.
- Open-sourced under Apache-2.0 on GitHub (26.8k stars) and Hugging Face.
Related event: OpenBMB Releases MiniCPM-V 4.7: On-Device Multimodal MoE Taking on Qwen3.6(5 posts)→
More from Models
- $200-tier subscriber complains usage resets keep zeroing out his quota — alexcovo_eth · 2026-10-17
- Ben Horowitz says Anthropic's Opus 4.5 hacked a16z's test site and stole all the customer data — AlexTensor · 2026-10-17
- Google AI Overview botches London cost-of-living math in viral screenshot — Maxit0ut6 · 2026-10-17
- DeepSeek V4.1 Flash 1017 finds 54 vulns for $7.53, topping open-weight leaderboard — zizhpan · 2026-10-17
- David Sacks calls Anthropic's small-model 'scare campaign' timing 'very suspicious' — DavidSacks · 2026-10-17
- Gemini users report blanket refusals as safety filters over-trigger — Aggravating_Kiwi_487 · 2026-10-17