---

base_model:

- openbmb/MiniCPM-V-4.7-1B-DSpark

license: apache-2.0

library_name: llama.cpp

tags:

- gguf

- dspark

- speculative-decoding

- magnitude

inference: false

---

MiniCPM-V-4.7-1B-DSpark GGUF

Q8_0 GGUF conversion of

openbmb/MiniCPM-V-4.7-1B-DSpark

for use with Magnitude.

This is a speculative drafter for

openbmb/MiniCPM-V-4.7-1B

and its tokenizer. It requires the corresponding target model and a runtime with DSpark support;

it is not a standalone language model. The draft has five layers and proposes 15 tokens per forward pass.

Artifact

• File: MiniCPM-V-4.7-1B-DSpark-Q8_0.gguf
• Quantization: Q8_0
• SHA-256: ec5c1d9fe4d8bb51d38c7097c7327b1f981efcf4ed44b72f7889ae255eb4c13b
• Source revision: e6bd0f240b7ef1f460783e3851983f6e8b3cfc50
• Target tokenizer revision: 6fde9ecee3e59c4b0ce555a9c774e179d46d19cc
• Conversion: F16 intermediate, then Q8_0, using llama.cpp revision 43fe9c64281ef735046adc025e9e7559a1f659a5 with the workshop's explicit drafter metadata handling.

Draft metadata

• dflash.attention.causal = false: the draft block uses bidirectional attention.
• dflash.sample_from_anchor = true: proposal selection starts at the anchor row.
• Causality is supplied explicitly for conversion from the DSpark reference implementation; the upstream checkpoint configuration omits it. Tokenizer dispatch uses the target's Qwen3.5 text tower.
• Tensor mappings and draft hidden states, Markov logits, confidence, and proposal selection were checked against reference code with shared synthetic target inputs. These are conversion checks, not end-to-end generation or performance measurements.