Recently, Liquid AI and Hugging Face officially released the DSpark draft model checkpoints for the three core models of the LFM2.5 series: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. By introducing a new speculative decoding path, this technology significantly improves the inference throughput without changing the model output quality at all.
Core Performance Improvements and Application Performance
In practical tests, the DSpark draft model showed impressive acceleration effects. On GPU, the overall throughput can be increased by up to 3.18 times; on edge devices, the maximum improvement can reach 2.87 times.
Especially in edge agent inference scenarios, the function call latency of LFM2.5-2.6B was reduced by an average of 57%. When using the M4Max MacBook Pro as the edge test platform, its output speed can reach up to 139 tokens per second, which greatly lowers the threshold for running agent applications locally, and the user experience even surpasses some proprietary cloud models.
How DSpark Works and Its Architecture Design
Traditional large language model inference is mainly limited by memory bandwidth, with most delays coming from streaming model weights from DRAM to SRAM. The emergence of speculative decoding aims to solve this pain point by using a lightweight draft model to quickly generate candidate tokens, which are then validated uniformly by the target model in a single forward pass, effectively distributing the cost of loading weights.
DSpark has deeply integrated existing methods, consisting of three core components:
- Parallel Backbone Network: Adopting a style similar to DFlash, it generates hidden states for all draft tokens uniformly in a single forward pass based on the context features of the target model.
- Sequential Head (Markov Head): By simulating Markov chains between adjacent tokens to increase dependencies, it effectively improves the acceptance rate of subsequent positions.
- Confidence Scheduling Validator: It predicts the survival probability of each token and automatically removes low-confidence suffixes when the validation cost exceeds the saved cost.
In terms of model training, the draft model uses a large-scale and diverse data mix including SFT, chat, code, and function calls. After strict ablation experiments, the initial version selected is an attention-only architecture with 5 layers and 9 blocks, with the total number of parameters controlled around 300 million.
Quality Alignment and Inference Ecosystem Support
Due to the characteristics of the speculative decoding mechanism, under greedy decoding, the draft token is only accepted when it perfectly matches the distribution of the target model. Once rejected, it is replaced by the token from the target model itself. Therefore, the generated output sequence maintains the same structure as the baseline greedy decoding, and there is no decrease in accuracy in any benchmark tests.
In terms of ecosystem support, DSpark achieved compatibility with mainstream inference frameworks on its release day:
- SGLang: Supports running on accelerators through dedicated integration and startup configurations.
- llama.cpp: Implements official build support and allows loading corresponding GGUF weights and draft model files via command line.
