LiquidAI Releases LFM2.5-DSpark Checkpoints for Faster Inference

LiquidAI released DSpark draft model checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. [1] This speculative decoding approach provides a large decoding speedup with only a minimal memory increase and no change in output quality. [1] It achieves up to a 3.18x throughput improvement on GPUs and up to 2.87x on-device. [1] For LFM2.5-2.6B, the method cuts function-calling latency by an average of 57%, advancing on-device agentic inference capabilities. [1] LFM-compatible DSpark integration is open-sourced upstream with day-one support for llama.cpp and SGLang. [1]

The DSpark architecture combines a DFlash-style parallel backbone, a lightweight sequential head modeled as a Markov chain, and a confidence-scheduled verifier. [1] The verifier predicts token survival probabilities and prunes low-confidence suffixes to save verification costs. [1] The initial draft models use simplified attention-only architectures with 5 layers and a block of 9. [1] They were trained for 15 epochs on a diverse data mix including SFT, chat, code, and function-calling data, selecting the epoch with the highest acceptance rate. [1]

LiquidAI's implementation of DSpark demonstrates that sophisticated speculative decoding techniques can substantially reduce latency for on-device and agentic tasks. By integrating a confidence-scheduled verifier alongside a DFlash-style backbone, the system addresses the traditional memory bottleneck of streaming weights from DRAM. Providing day-one support for popular frameworks like llama.cpp ensures immediate usability for developers building local and edge AI applications. What the sources don't address: How the draft models perform across varying batch sizes during high-concurrency enterprise deployments.