Inference up to 3.2× faster with LFM2.5‑DSpark
agents gpu huggingface inference llama open-source
| Source: Hugging Face | Original article
LFM2.5-DSpark boosts inference speed by up to 3.2×, delivering about 3.18× higher GPU throughput and nearly 2.9× faster on‑device performance.
Liquid AI announced the release of “LFM2.5‑DSpark,” a set of speculative‑decoding draft checkpoints that promise markedly faster inference for its LFM2.5 model family. The company made the draft checkpoints available for three models – LFM2.5‑1.2B‑Instruct, LFM2.5‑2.6B and the mixture‑of‑experts LFM2.5‑8B‑A1B – in both Safetensors and GGUF formats. Benchmarks show up to a 3.18‑fold throughput boost on a single Nvidia H100 GPU and up to a 2.87‑fold increase on Apple‑silicon Macs. In addition, the draft models cut function‑calling latency by roughly 57 % on average, moving the LFM2.5 line closer to practical on‑device, agentic AI.
The announcement matters because inference speed remains a bottleneck as large language models proliferate across cloud and edge environments. By pairing a lightweight draft model with the full‑size LFM2.5 checkpoint, DSpark delivers higher token‑per‑second rates without altering final token distributions, effectively squeezing more work out of existing hardware. The open‑source integration with llama.cpp and SGLang means developers can adopt the acceleration with minimal code changes, echoing recent moves by other players to tighten the inference stack – from Etched’s AI‑inference chips to OpenAI’s “Ultrafast” API tier and the surge in eSSD‑based server deployments.
What to watch next includes early adoption metrics from the developer community, especially on mobile and edge devices where the 57 % latency cut could enable fully offline assistants. Further performance validation on other accelerators, as well as potential integration into managed inference services such as Google Cloud’s Gemini Enterprise Agent Platform, will indicate how quickly DSpark reshapes deployment economics. If the gains hold up at scale, Liquid AI’s approach could become a reference point for future speculative‑decoding toolkits.
Sources
Back to AIPULSEN