LiquidAI has unveiled the DSpark draft model checkpoints for three models within its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. This innovative approach introduces a speculative decoding path that enhances decoding speed while maintaining output quality, achieving up to 3.18x throughput improvement on GPUs and 2.87x on-device.
The DSpark model aims to facilitate on-device agentic inference by reducing function-calling latency by an average of 57% for the LFM2.5-2.6B model. Furthermore, it offers day-one support for llama.cpp and SGLang, with LFM-compatible DSpark integration available as open-source.
Mechanics of DSpark
The traditional decode phase in large language model (LLM) inference is often constrained by memory bandwidth. Most latency arises from transferring weights from DRAM to SRAM, rather than from computational intensity. Speculative decoding addresses this issue by employing a lightweight draft model to generate candidate tokens, which are then verified by the target model in a single forward pass. This method allows for shared weight loading costs across all tokens being verified.
DSpark combines several components: a DFlash-style parallel backbone conditioned on the target model’s context features, a lightweight sequential head modeled as a Markov chain to enhance inter-token dependency, and a confidence-scheduled verifier that predicts token survival probabilities, pruning low-confidence suffixes when verification costs outweigh benefits.
Training and Architecture
The DSpark models were trained using a diverse dataset that includes supervised fine-tuning (SFT), chat, code, and function-calling data. The initial draft models are attention-only, featuring 5 layers and approximately 300M parameters. Each model underwent 15 epochs of training, selecting the epoch with the highest acceptance rate for deployment.
Performance Metrics
In terms of inference speed, the DSpark draft models exhibit substantial throughput improvements across various configurations. For instance, the LFM2.5-2.6B model achieves around 140 tok/s on a MacBook Pro, significantly enhancing user interactivity compared to many proprietary cloud models. The models also maintain benchmark accuracy, ensuring that the emitted sequences remain consistent with baseline greedy decoding.
The DSpark draft model checkpoints are now available on Hugging Face in both Safetensors and GGUF formats, inviting developers to explore their capabilities further.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
