MiniCPM-V 4.7 35B-A3B added to BenchGecko · open-weight multimodal MoE
OpenBMB released MiniCPM-V 4.7 on 2026-10-12. BenchGecko lists the 35B-A3B variant, an Apache-2.0 multimodal MoE with a 262,144 token position limit.
What landed
MiniCPM-V 4.7 35B-A3B is the newest model in OpenBMB's MiniCPM-V series, released on 2026-10-12. The series now ships in two sizes, 35B-A3B and 1B. The 35B-A3B variant is the first Mixture-of-Experts model in the series and the one BenchGecko lists.
Specs
From the Hugging Face model card and config:
- Parameters · 35B total, about 3B active per token
- Experts · 256, with 8 routed per token
- Layers · 40, mixing linear attention with full attention every fourth layer
- Position limit · 262,144 tokens (max_position_embeddings in config)
- Inputs · text, single image, multiple images, video (up to 128 sampled frames by default)
- Output · text
- Visual token compression · mixed 4x and 16x modes
- Backbone · Qwen3.5 and Qwen3.6 language models with a SigLIP2 vision module
BenchGecko stores the context field as 262,144 from the published config. OpenBMB does not publish a separate serving limit.
Licence and formats
Weights are open under Apache-2.0. OpenBMB publishes FP8, int4, AWQ, GPTQ and GGUF variants plus a DSpark draft model for speculative decoding. Supported runtimes include SGLang, vLLM, llama.cpp and Ollama, with fine-tuning through SWIFT and LLaMA-Factory.
On-device claims
OpenBMB reports a decoding speed of 13+ tokens per second using about 3 GB of memory on an iPhone 17 Pro Max. The method combines Lookahead Expert Scheduling, which preloads experts predicted for the next layer, with layer-wise quantization and DSpark speculative sampling. These are vendor figures, not BenchGecko tests.
Pricing
No first-party API price is published, so BenchGecko leaves the pricing fields empty.
Sources
- MiniCPM-V 4.7 35B-A3B model card
- MiniCPM-V GitHub repository
- BenchGecko model page: /model/minicpm-v-4-7-35b-a3b
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.