Liquid AI releases DSpark draft models for LFM2.5
The checkpoints add speculative decoding for three LFM2.5 models with llama.cpp and SGLang support.
Why it matters
Speculative decoding is becoming a practical route to lower-latency local and server inference without changing the target model. Liquid AI’s release gives developers an open implementation path for testing those gains on both GPU and MacBook-class hardware.
The key points
- 1.DSpark checkpoints cover three LFM2.5 models.
- 2.Liquid AI reports up to 3.18x H100 throughput improvement.
- 3.Safetensors and GGUF versions are available on Hugging Face.
Liquid AI released DSpark draft model checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. The company says the checkpoints add a speculative decoding path that improves inference throughput by up to 3.18x on an H100 and up to 2.87x on-device, while preserving greedy-decoding output quality. The draft models are available on Hugging Face in Safetensors and GGUF formats, with support for llama.cpp and SGLang.
⚡ Try this today
If you run LFM2.5 models, test the matching DSpark checkpoint in llama.cpp or SGLang against your own latency and output-quality baselines.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Models
LiquidAI releases QAD Q4_0 checkpoints for LFM2.5
The 4-bit GGUF checkpoints target edge deployment with reduced quantization quality loss.
Artificial Analysis benchmarks GLM-5.3
The GLM-5.3 benchmark page drew discussion on Hacker News and r/LocalLLaMA.
Qwen3.8 27B scores 52 on Artificial Analysis
Reddit and Hacker News users flagged a reported benchmark jump for Qwen3.8 27B.