OpenBMB released MiniCPM-V-4.7-1B-DSpark, a speculative decoding draft model for MiniCPM-V-4.7-1B
OpenBMB released MiniCPM-V-4.7-1B-DSpark on Hugging Face, a 5-layer draft model with 213,547,777 parameters trained for exact pairing with MiniCPM-V-4.7-1B and its tokenizer. It is licensed Apache-2.0 and supports vLLM. The source release was published on 2026-10-08.
China context
- Original name
- 面壁智能
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the Apache-2.0 licensed weights from Hugging Face and integrate the draft model with vLLM for speculative decoding of MiniCPM-V-4.7-1B.
- For investors
- The release of a dedicated speculative decoding component for MiniCPM-V-4.7-1B indicates ModelBest's focus on inference efficiency for edge and on-device deployment, which may strengthen its position in the open-weights vision-language model market.
OpenBMB released MiniCPM-V-4.7-1B-DSpark, a DSpark draft checkpoint designed for speculative decoding with MiniCPM-V-4.7-1B. The draft model has 5 layers, 213,547,777 parameters, and generates 15 draft tokens per forward pass, targeting hidden-state layers [3, 7, 11, 15, 19] of the target model. It was trained on 2M sequences for 6 epochs using AdamW with a combined CE + L1 + confidence loss objective. Reported acceptance lengths range from 3.370 (General text) to 5.891 (Math text), with vision domains such as MMMU at 3.452 and OCRBench at 4.752. The model is released under Apache-2.0 and can be used with vLLM.
The draft model uses 15 draft tokens per forward pass and targets five specific hidden-state layers of MiniCPM-V-4.7-1B, enabling speculative decoding that reduces inference latency. Reported acceptance lengths vary by domain, with text-based tasks (Math 5.891, Code 5.347) showing higher acceptance than general text (3.370) and vision tasks (MMMU 3.452).
Developers deploying MiniCPM-V-4.7-1B on resource-constrained devices can reduce inference latency by pairing it with this draft model, which is available under Apache-2.0 and supports vLLM. The release extends the MiniCPM-V ecosystem with a dedicated speculative decoding component, potentially lowering the cost of serving the 1B vision-language model.
For developers using MiniCPM-V-4.7-1B, this draft model offers a path to faster inference without changing the target model, potentially reducing serving costs and improving responsiveness for on-device or edge deployments. The Apache-2.0 license permits commercial use.
A verifiable next signal is whether OpenBMB releases DSpark draft checkpoints for other MiniCPM-V variants or integrates the draft model into official inference examples beyond vLLM. Independent benchmarks of end-to-end speedup on target hardware would confirm the practical benefit of the reported acceptance lengths.