OpenBMB releases MiniCPM-V-4.7-1B-DSpark draft model for speculative decoding
AIOpenBMB has released MiniCPM-V-4.7-1B-DSpark, a 213,547,777-parameter draft checkpoint that pairs with the MiniCPM-V-4.7-1B target model and its tokenizer. The draft model predicts 15 tokens per forward pass and reports acceptance lengths from 3.370 on general tasks to 5.891 on math, with vLLM usage shown via the dspark speculative method.
Source details
OpenBMB (MiniCPM) · new models on Hugging Face · huggingface.coPublished · added here
Original title: openbmb/MiniCPM-V-4.7-1B-DSpark