# OpenBMB releases GGUF‑quantized MiniCPM‑V draft model on Hugging Face

> A 35‑billion‑parameter multimodal model draft (DSpark) is now available in GGUF format for llama.cpp and vLLM.

Oossa · 2026-10-14 · https://oossa.com/en/openbmb-releases-gguf-quantized-minicpm-v-draft-model-on-hugging-face

OpenBMB has added a GGUF‑quantized version of MiniCPM‑V‑4.7‑35B‑A3B‑DSpark to Hugging Face. The draft checkpoint works with the matching MiniCPM‑V target model for speculative decoding, allowing up to 15 draft tokens per forward pass. It can be run with llama.cpp or vLLM using the provided command‑line flags. The model supports vision‑text inputs and runs with BF16 precision in the original weights.

## The facts

- Model size: 35 billion parameters, 389 M draft parameters
- Speculative draft tokens per pass: 15

## Why it matters

Developers can now run a large multimodal model faster on consumer hardware by using the lightweight GGUF draft.

## Sources & references

1. [New model on Hugging Face: openbmb/MiniCPM-V-4.7-35B-A3B-DSpark-gguf (text-generation)](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B-DSpark-gguf) – OpenBMB, 2026-10-14

Last updated: 2026-10-14