MiniCPM-V 4.7 beats Qwen3.6 on doc understanding, +14 pts on OmniDocBench
NielsRogge · x · 2026-10-17
Niels Rogge shares official benchmark numbers for MiniCPM-V 4.7 (35B-A3B with 4x visual token compression) against Qwen3.6-35B-A3B. With identical 35B-total/3B-active MoE footprints:
- OmniDocBench v1.6: 94.2 vs 80.2 (+14.0 pp)
- CharXiv reasoning: 75.6 vs 67.5 (+8.1 pp)
- OCRBench: 91.2 vs 86.0 (+5.2 pp)
Since parameter counts match, the gains are attributable to the 4x compression method itself, showing strong document and chart understanding.
Related event: OpenBMB Releases MiniCPM-V 4.7: On-Device Multimodal MoE Taking on Qwen3.6(5 posts)→
More from Models
- DeepSWE benchmark skeptics told to wait — Datacurve points to fully public repo — scaling01 · 2026-10-18
- SubQ Snap launches: 384K context, sub-second 100K-token decisions at $0.05/M tokens — eyishazyer · 2026-10-18
- mrfakename uploads GGUF quants of DeepSeek-V4.1-Flash-1017 to Hugging Face — realmrfakename · 2026-10-18
- DeepSeek-V4.1-Flash-1017 GGUF quants appear on Hugging Face with 473GiB Q8_0 — realmrfakename · 2026-10-18
- Power User's Model Split: 80% Astra, 10% Opus 5.5, 10% Argon 4 for Writing — soumitrashukla9 · 2026-10-18
- Pangram AI detector flags human-written paragraphs, sparking author fears — ArkhamDreamerZero · 2026-10-18