769 tokens per second on a single NVIDIA RTX 4090. Spark-X2.6-1.7B with DSpark reaches that peak in the published GSM8K benchmark using a Q4_K_M target with llama.cpp, up from 424 tok/s without DSpark.
For BF16, the highest reported speedup is 2.43x, also on GSM8K and RTX 4090: 220 → 536 tok/s. The throughput peak and speedup peak use different target precisions.
❤️ Give the model a Like and download the weights: Hugging Face · ModelScope.
A 345M-parameter BF16 draft model proposes up to seven tokens per pass. Spark-X2.6-1.7B verifies those candidates in parallel, reducing sequential target decoding steps. The draft uses five sliding-window attention layers and is provided in BF16 GGUF format. Pair it with the Spark-X2.6-1.7B target model, using BF16 or quantized target weights.
Across AIME25, GSM8K, MATH500, HumanEval, MBPP and Arena-Hard-v2, the model card reports:
| Target / device / runtime | Average speedup | Average throughput, without → with DSpark |
|---|---|---|
| BF16 · RTX 4090 · llama.cpp | 2.10x | 220 → 463 tok/s |
| BF16 · M5 Pro · MLX | 1.88x | 68 → 128 tok/s |
| Q4_K_M · RTX 4090 · llama.cpp | 1.53x | 424 → 650 tok/s |
| INT4 · M5 Pro · MLX | 1.20x | 138 → 165 tok/s |
The acceptance-length evaluation uses seven draft tokens and a maximum output length of 4,096 tokens. Average acceptance lengths range from 3.5763 to 4.3158 across the four reported sampling/thinking configurations. Acceptance length includes an additional target token, giving a maximum of eight.
These are the published model-card results. Performance depends on the device, target precision and workload; the poster and model card show the full tables.
To run it locally, follow the model card setup for XHToken/llama.cpp. The documented commands target Linux, Bash and NVIDIA GPUs. Integration into upstream llama.cpp is in progress; the card reports MLX performance separately. Spark-X2.6 uses the Apache 2.0 license.
Like the model. Download the magic. Give it a Like on Hugging Face and ModelScope, download the draft and target weights, and share your device, runtime, precision and measured results with the community.
Download the full-resolution English poster — 7920 × 15400 PNG
Top comments (0)