OpenBMB's DSpark draft head hits 7.4 acceptance length for MiniCPM-V 4.7
NielsRogge · x · 2026-10-17
Released alongside MiniCPM-V 4.7, DSpark is a speculative decoding draft checkpoint trained for exact pairing with the 35B-A3B target model (1B variant too):
- Draft side: 5 layers, 389M params, proposes 15 tokens per forward pass for main-model verification
- Trained on 2M vision+text sequences generated by the target model, 6 epochs, CE + L1 + confidence loss
- Acceptance lengths: Code 7.39, Math 6.69, MathVista 6.13, CharXiv 5.48, General 3.50
- Keeps V4.6's 4x/16x visual token compression, so images cost fewer tokens
- GGUF and vLLM configs (with speculative-config example) available, Apache-2.0
Related event: OpenBMB Releases MiniCPM-V 4.7: On-Device Multimodal MoE Taking on Qwen3.6(5 posts)→
More from Infra
- DeepSeek-V4.1-Flash-1017 GGUF quants appear on Hugging Face with 473GiB Q8_0 — realmrfakename · 2026-10-18
- DeepSeek-4.1-Flash Runs on 4x RTX Pro at 300+ TPS With NVMe Offload — TheZachMueller · 2026-10-18
- SGLang Summit sponsors jump from 11 to 36 in one week, 2,000+ attendees expected — ying11231 · 2026-10-18
- Agent search drops matrix multiplication exponent below 2.7 at ~500M matrix size — IgorCarron · 2026-10-18
- Dell says memory chips are the real AI server bottleneck — and it'll 'get worse' next year — firstadopter · 2026-10-18
- NVIDIA App reserves 32GB of RAM by default on 64GB machines, hurting local AI — draginol · 2026-10-18