Spark-X2.6-4B DSpark Draft
This model is part of the validation set for Draftfit, a platform for building, converting and evaluating speculative decoding drafts with shared target features.
A DSpark draft for Spark-X2.6-4B. It requires the target model for text generation.
Acceptance
| Evaluation | Official draft | This draft |
|---|---|---|
| Development | 37.93% | 40.11% |
| Confirmation | 35.85% | 38.44% |
Measured with the same greedy PyTorch prefix verifier, seven draft proposals, 64 output tokens and 24 prompts per partition. Acceptance is accepted/proposed draft tokens, excluding target bonus tokens. All outputs match the target exactly.
These small internal benchmarks had prior evaluation exposure. Both paired prompt bootstrap 95% intervals include zero. Production serving speed and full original C++ runtime parity are unverified.
Use
Load model.safetensors into the included SparkDSpark(official_config()) reference with safetensors.torch.load_file. A compatible Draftfit Spark target adapter is required; provenance is in experiment-manifest.json. This is a custom reference checkpoint, not a drop-in vLLM, SGLang or MLX model.